Forecasting short term yield changes using order �ow

Siri Valseth

September 15, 2009

Abstract

This study investigates the predictive power of bond market order �ow on yield changes

while controlling for traditional term structure variables. Both in-sample and out-of-sample

predictions of daily and monthly yield changes based on lagged order �ow are compared to

predictions based on traditional term structure models and the random walk. The results

show that models including order �ow clearly outperforms models based on forward rates and

historical rates. Using a unique data set including dealer identities the results further indicate

that dealers are heterogeneous. Varying predictive power of the order �ow of di¤erent groups

of dealers imply that dealers have di¤erent sources of information. Dealers with a high share

of delayed publication trades appear to have stronger predictive power than other dealers. This

suggests that by hiding trades, dealers can utilize the private information in these, and do

pro�table trades in the interdealer market before other dealers have the chance to update their

beliefs.

1 Introduction

Forecasts of yield changes are of interest to investors, fund managers, analysts and monetary

policy makers, but it is di¢ cult to �nd a model that produces good forecasts. The classical

expectations hypothesis, which dominates the literature on interest rate predictability, implies

that forward rates can predict interest rate changes. However, numerous empirical studies �nd

that forward rates are poor predictors of interest rates, especially over short horizons. Fama

and Bliss (1987) and Campbell and Shiller (1991) �nd that forward rates and yield spreads have

little predictive power for future interest rates and conclude that their �ndings are inconsistent

with the expectations theory. These results have supported the view that interest rates follow

a "random walk", implying that today�s interest rate is the best predictor for future interest

rates. Recently, a micro-based approach has appeared as a promising venue to asset price

predictability. This market microstructure approach uses order �ow as a predictor variable.

Order �ow is a measure of the net buying pressure in the market and is here de�ned as the

number of buyer initiated trades minus the number of seller initiated trades during a day. Evans

and Lyons (2005) use foreign exchange market order �ow to predict exchange rates and �nd

that the out-of-sample predictive power of order �ow clearly outperforms the random walk. In

this paper the predictive power of bond market order �ow is compared to the predictive power

of forward rates and to the random walk.

The vast literature on interest rate predictability is based on the traditional "e¢ cient mar-

ket" assumption that asset prices completely and instantaneously re�ect all relevant informa-

tion. In the market microstructure literature, the price formation process is assumed to take

time. According to this literature, new information is imbedded into asset prices through two

channels; a direct channel and an indirect channel. Whereas the direct channel instantaneously

incorporates public information into prices, the indirect channel gradually incorporates private

information into prices through order �ow. Private information is de�ned by Lyons (2001) as in-

formation not known by all people that produces a better price forecast than public information

alone. Private information may thus include stricktly private information and heterogeneous

interpretations of public information.

Price formation through the indirect channel is referred to as price discovery. According to

O�Hara (2003) "markets have two important functions - to provide liquidity and price discovery-

and these functions are important for asset pricing". Price discovery describes the process in

which asset prices adjust to full information prices. Market participants observe the trade �ow,

infer private information from the net order �ow and set prices accordingly. By observing the

order �ow market participants will gradually infer good news when net order �ow is positive

and bad news when net order �ow is negative. The source of private information in order �ow

may be stricktly private based on a dealer-customer relationship. For example, a dealer could

through a customer trade obtain information about the unwinding of a large hedging position in

bonds, leading her to sell bonds because she expects lower prices and higher yields in the future.

The source of private information may also be an interpretation of public information from a

2

dealer with superior analytical skills. For example, a dealer analyzes local retail sales which are

showing decreasing sales numbers, and she expects weak national retail sales next week. She will

thus buy short bonds ahead of the publication next week because she expects the central bank

to reduce the key interest rate in order to prevent an economic downturn. Private information

will be re�ected in the trades of these dealers, but since the information is not picked up by the

majority of market participants at once, it will not be re�ected instantaneously in the current

yield curve. Instead other traders will observe the order �ow and infer information from this.

Eventually new private information will be impounded into the yield curve.

The main contribution of this paper is to investigate whether a market microstructure

perspective can improve the predictability of yield changes while controlling for traditional

term structure variables. The paper is the �rst to include bond market order �ow in term

structure models and compare in-sample and out-of-sample predictions based on order �ow

with predictions based on forward rates and the random walk. Another important contribution

is to test for dealer heterogeneity by including individual dealer order �ow. Through a unique

data set including the identities of the buying and the selling dealers the existence of information

asymmetries among the dealers can be tested. The data set includes di¤erent types of trades and

the study also addresses the issue of market transparency by revealing the importance of hidden

trades for the predictive power of order �ow. These contributions increase our knowledge about

the type information imbedded in order �ow. The paper also contributes by making short term

forecasts at the daily horizon, by discussing the link between predictability of yield changes and

excess bond returns and �nally by controlling for the persistence in monthly observations based

on overlapping data by employing non-overlapping monthly data for the monthly forecasts.

This paper documents that models including order �ow clearly outperforms the models

based on forward rates only and the random walk. Traditional term structure variables rep-

resented by the Fama-Bliss forward spread and �rst three principal components of all forward

rates, have litte explanatory power at the daily horizon. At the monthly horizon the third

principal component is economically and statsistically signi�cant. This component appear to

be capturing the uniqueness of forward rates as the correlation with the corresponing factor

decomposition based on zero-coupon rates is much smaller than for the other principal compo-

nents at only 18 percent. Lagged bond market order �ow can forecast changes in bond yields

of all maturities at both daily and monthly horizons. Medium term order �ow appears to have

the strongest predictive power for all yield changes at the daily horizon except for the 10-year

yield, where only long term order �ow is signi�cant.

At the monthly horizon, only order �ow based on the most liquid bonds, which are bonds

with a remaining time to maturity of between 1 and 7 years, is included. Since principal

components analysis show that 98 percent of the variance is due to parallell shifts in the yield

curve, the short and medium term order �ow are added to one order �ow variable. This

monthly variable can explain up to 18 percent of the variation in yields in the following month.

At the daily horizon the out-of-sample predictions based on order �ow gives a ratio of the mean

3

squared error over the mean sqared error of the random walk of 0.98 and 0.99, whereas the

out-of-sample predictions based on the forward rate factors gives a ratio of 1.01. At the monthly

horizon the out-of-sample predictions based on order �ow gives a ratio of the mean squared

error over the mean sqared error of the random walk of between 0.84 and 0.93 for the di¤erent

yield maturities, whereas the out-of-sample predictions based on the third forward rate factor

gives a ratio of between 0.95 and 1.00. At the monthly horizon models including both order

�ow and the third principal component of forward rates acheive a ratio of mean squared errors

as low as 0.80 when compared to the random walk model. These results strongly indicate that

inclusion of market microstructure variables in traditional term structure models will improve

the forecasts of future yield changes substantially.

The results further show that bond dealers are heterogeneous. The predictive power of the

order �ow of dealers with a high share of customer trades with delayed publication is signif-

icantly higher than that of dealers with a low share of delayed publication trades. Customer

trades with delayed publication are trades entered in the trading system by the dealer with a

�ag indicating that they are hidden from the other market participants for a certain time. This

time lag has changed over the sample period. Until mid-2002 there was a two hour delay, after

that time the delay has been until the end of the trading day at 4 pm. This delay is meant to

allow market makers to unwind their inventory positions at minimal cost. The higher predic-

tive power of dealers that have a higher ratio of delayed publication trades indicate that these

dealers either have more informed customers than other dealers or that they have a di¤erent

strategy/ability than other dealers in order to bene�t from private information in customer

trades. If dealers have more informed customers, are able to identify them and use the �ag for

delayed publication on these trades, they will gain private information that will be hidden from

the other market participants, they can trade on this information in the interdealer market

before prices are updated with this information. When these trades become visible to the other

dealers, prices will be updated according to the net order �ow of these trades. Also, if dealers

have the same share of informed customers, but choose di¤erent strategies, their order �ow can

have di¤erent predictive power. The ones who choose the strategy of hiding the trades from

the others, will be able to trade in the interdealer market using the information in the customer

trade before the information is known by the others and thus preventing the others to adjust

their prices accordingly. The ones who choose not to hide customer trades, independent of

whether the customer is assumed to be informed or not, will give the other dealers a chance of

updating their beliefs and hence update their prices, and can to a much lesser extent bene�t

from any private information in own customer trades.

A buying pressure expressed by a positive net order �ow predicts a decrease in yields and

an increase in bond prices. As much of the term structure literature focus on the predictability

of bond risk premia, this paper reports in-sample and out-of-sample results of the predictive

power of aggregate order �ow on both yield changes and excess bond returns. In line with

previous research (Fama and Bliss (1987) and Cochrane and Piazzesi (2005)) the results here

show that forward rates are better predictors of excess bond returns than of yield changes.

4

However, this is not the case for order �ow, which appear to have similar predictive power for

both excess returns and yield changes. Thus the private information contained in order �ow

may be related to risk premia as well as to future short interest rates.

Recent studies have shown that long horizon forecasts based on overlapping observations of

highly persistent variables may lead to spurious results. Boudoukh, Richardson and Whitelaw

(2006) show that under the null hypothesis of no predictability, many persistent variables

produce coe¢ cient estimates and R2 �s that are highly correlated across horizons. In order to

correct for a possible bias due to the high persistence of monthly order �ow based on overlapping

observations, this study uses non-overlapping monthly observations in the predictive regressions.

The results document that order �ow has predictive power for yield changes at the monthly

horizon.

The rest of this paper is organized as follows. Section 1.1 gives a brief overview of the

relevant literature. Section 1.2 describes the data set and trading conventions in the Norwegian

government bond market. Section 1.3 presents the models and econometric framework. Sections

1.4 and 1.5 reports the in-sample and out-of-sample results respectively. Section 1.6 discusses

the predictive power of the order �ow of di¤erent groups of dealers. Section 1.7 summarizes

and concludes.

1.1 Related literature

This paper is related to three segments of the �nance literature, the term structure literature,

the market microstructure literature and the literature on asset return predictability. The

vast literature on the term structure of interest rates is based on the expectations hypothesis.

The classical expectations hypothesis states that bond yields are expected values of average

future short term rates, and implies that forward rates are expected future spot rates and thus

can predict future yield changes. It also states that the holding period return on bonds of

all maturities should be equivalent. The expectations hypothesis can be modi�ed to include

constant risk premia, implying a one-to-one relationship between forward rates and expected

future interest rates. However, for the past twenty years empirical studies have produced

evidence against the expectations hypothesis, indicating that yields contain time-varying risk

premia and that these premia, de�ned as the expected bond price minus the forward price or

the expected exess return on bonds, are predictable.

Fama and Bliss (1987) and Campbell and Schiller (1991) investigate whether forward rates

and yield spreads can predict future changes in interest rates, but �nd little evidence in favour

of this. Fama and Bliss (1987) use the forward spread, de�ned as the 1-4 year ahead forward 1-

year rate minus the current 1-year rate, to predict 1-year yield changes and �nd that the forward

spread is a poor predictor of 1-year rates, especially at short horizons. They �nd instead that

the forward spread predicts 1-year excess returns on bonds and conclude that the bond risk

5

premium varies over time and is predictable. Campbell and Schiller (1991) study whether the

yield spread, measured as the di¤erence between yields of maturities up to ten years and the

short term interest rate, can predict interest rates of all maturities over several horizons. They

�nd that a high yield spread predicts a fall in long yields, which is counter to the expectations

hypothesis. However, they �nd that a high yield spread also predicts an increase in short rates,

which is in accordance to the theory. Cochrane and Piazzesi (2005) strenghten the evidence

againt the expectations hypothesis by showing that a linear combination of all forward rates

can predict bond risk premia on 1- year horizons with a substantially higher forecasting power

than the maturity speci�c forward spread. All studies use monthly observations, Fama and

Bliss (1987) for the period 1964-1985, Campbell and Schiller (1991) for the period 1952-1987

and Cochrane and Piazzesi (2005) use the Fama-Bliss data updated through 2002 to predict one

year excess returns. This study builds on the papers mentioned above by incorporating forward

rates in the predictive regressions. It di¤ers by adding order �ow as a predictor variable and

by using daily data over forecasting horizons of one day and one month, which is substantially

shorter than in earlier studies. In addition to shorter forecast horizons, this study performs

out-of-sample forecasting which should be of interest for analysts and investors with short

investment horizons.

Kessler and Scherer (2009) extend the Cochrane and Piazzesi (2005) model by applying it

to international bond markets, and con�rm that the model apply to other markets than the

Treasury market. They �nd that forward rates predict bond excess returns in seven major

bond markets. Engsted and Tanggaard (1995) predict short and long term Danish interest

rates using yield spreads for the period 1976-1991, but �nd that the predictive power of yield

spreads disappears under recent monetary policy regimes. They also �nd that the yield spread

predicts long rates in a direction opposite to that implied by the expectations hypothesis. This

study adds to the literature on international bond markets by using data from the Norwegian

government bond market.

Recent studies highlight the importance of using information beyond that contained in the

yield curve to convey the cyclical pattern in bond market risk premia. Several studies �nd that

factors other than forward rates and yield spreads have predictive power for bond risk premia.

Ludvigson and Ng (2008) �nd that macroeconomic fundamentals can forecast variation in

bond excess returns. Principal components of a large set of macroeconomic indicators convey

the cyclical pattern in bond market risk premia. Ilmanen (1995) shows that �nancial market

variables can forecast excess government bond returns in six countries. He concludes that

wealth dependent relative risk aversion appears to be an important source of bond return

predictability. Cooper and Priestly (2008) document that the output gap has predictive power

for both stock excess returns and bond excess returns. This study also uses information beyond

the current yield curve to predict bond risk premia. However, instead of macroeconomic and

�nancial market indicators that are theoretically related to bond market risk, the variable used

here is of a di¤erent nature. Order �ow re�ects private information that is not yet incorporated

into the yield curve. This information may include both fundamental and non-fundamental

6

information held by market participants, and there could therefore be related to risk premia as

well as expected future short rates.

This paper is also related to the rapidly increasing literature on market microstructure.

There are by now quite a few studies on bond markets. Brandt and Kavajecz (2004) examine

the price formation process in the US Treasury market based on interdealer order �ow. They

�nd that up to 26 percent of daily yield changes can be accounted for by net interdealer order

�ow. The yield changes induced by order �ow are found to be permanent, excluding inventory

concerns to be the cause of the yield changes. Green (2004) studies the impact of trading on

government bond prices surrounding the release of macroeconomic news. He examines trades

that take place in the half hour before and the half hour after the release of a macroeconomic

announcement. He �nds that the informational role of trading increases after macroeconomic

announcements, suggesting that the release of public information increases the information

asymmetry in the market. Pasquariello and Vega (2007) show that unanticipated order �ow in

US treasuries 1 has signi�cant and permanet impact on daily bond yield changes on both news

days and no-news days. Valseth (2008) �nds similar results from the Norwegian government

bond market where interdealer order �ow explains more than a quarter of daily yield changes

including all news and no-news days. She also �nds that aggregate interdealer order �ow is more

informative than aggregate customer order �ow. Underwood (2008) studies the cross-market

information content of stock and bond order �ow and �nds that they play an important role in

explaining cross-market returns. This study di¤ers from the above mentioned by focusing on

out-of-sample predictions of daily and monthly yield changes.

Finally, this paper is related to the extensive literature on predictability of asset returns.

Goyal and Welch (2008) reexamines the performance of variables that have been suggested in

earlier studies to be good predictors of the equity premium and �nd that most of the models have

predicted poorly over the last 30 years both in-sample and out-of-sample. They emphasize the

importance of testing for genuine superior and stable in-sample and out-of-sample performance

in years after the model identi�cation in order to verify the predictive variables. Boudoukh,

Richardson and Whitelaw (2005) show that for persistent regressors, the estimators are almost

perfectly correlated across horizons under the null hypothesis of no predictability. Common

sampling errors across equations leads to OLS coe¢ cient estimates and R2 that are roughly

proportional to the horizon under the null hypothesis implying that evidence on predictability

based on overlapping data may well be spurious. They conclude that the key determinants of

long-horizon predictability based on overlapping observations are the extent of predictability at

short horizons and the persistence of the regressor, and recommend researchers to be cautious

when interpreting long horizon forecasts based on persistent predictors. This paper seeks to

1They de�ne unanticipated order �ow as the order�ow over thirty minute intervals that are not explaned bylagged thirty minute order �ow or thirthy minute quote revision. They use 19 lags which equals a trading day.Unaticipated order �ow is calculated by adding up the 19 error terms within each day.

7

test whether order �ow is a genuine predictor by employing some of the controls suggested in

this literature.

Predictability based on microstructure models are documented in recent studies. Evans

and Lyons (2005) document that models including order �ow in foreign exchange markets have

signi�cant out-of-sample forecasting power for exchange rates. They compare 4 models that

forecast exchanges rates over horizons from one day to one month, where two are based on

macroeconomic factors and two are based on aggregate and disaggregated customer order �ow

data, with the random walk. They show that out-of-sample forecasts of exchange rate changes

based on lagged order �ow consistently outperform both the random walk and models based on

macroeconomic variables. They �nd that while the macroeconomic models are outperformed

by the random walk, the microstructure model with disaggregated order �ow consistently beats

the random walk. The forecasts from their model account for nearly 16 percent of the sample

variance in monthly exchanges rates. They use overlapping daily data for forecasting horizons

exceeding one day. Evans and Lyons (2008) show that order �ow forecast both the exchange rate

and its underlying macroeconomic determinants. They conclude that the forecasting relation-

ship arises because the same macro information revealed in order �ow is useful for determining

the foreign exchange risk premium, generating rational forecastability in returns.

This study is related to Evans and Lyons (2005), but di¤ers in several ways in addition to

adressing a di¤erent asset market. It is similar in comparing the predictability of lagged order

�ow with traditional models and the random walk, but di¤erent in using interdealer order �ow

instead of customer order �ow and in employing a market data set including dealer identities.

This paper also di¤ers in that it attempts to reveal whether the information contained in order

�ow is related to bond risk premia by studying the predictability of both yield changes and

excess returns. While Evans and Lyons (2005) do not test the strong predictability on the

one month horizon for possible bias due to the persistence of overlapping observations, this

paper adresses the potential problem by employing both non-overlapping monthly data and

overlapping daily observations on the monthly forecasting horizon.

1.2 Data and trading environment

The analyses in this paper are based on a unique data set on bond trades in the Norwegian

Government bond market. The data include all transactions and best bid and ask submitted

by members of Oslo Stock Exchange (OSE) in the six year period from September 6, 1999 to

September 30, 2005. Each transaction includes date, time, price, amount, identity of buying

and selling dealer and type of trade. Repo trades and trades with a trade amount of less than 1

million NOK (180 000 USD) are excluded from the sample as the informational content of these

trades is assumed to be limited. The number of trades used in the analysis is 66650 and in-

cludes ordinary over-the-counter trades, over-the-counter trades with non-standard settlement,

automatch (electronic) trades, delayed publication trades and trades registered outside market

8

opening hours. Trades between di¤erent buying and selling dealers are de�ned as interdealer

trades, and trades with the same dealer as both the buying and selling dealer are de�ned as

customer trades. Evans and Lyons (2005) base their analysis on customer order �ow. They �nd

that the model based on the order �ow of six di¤erent groups of end-users have a signi�cant

higher predictive ability than the model based on aggregate end-user order �ow. This study

uses aggregate interdealer order �ow and �nd that forecasts based on these have a predictive

power similar to the disaggregated model of Evans and Lyons (2005). Only interdeater order

�ow is included in the forecasting models as Valseth (2008) documents that the information

content in aggregate interdealer order �ow is substantially higher than the information content

in aggregate customer order �ow.

The order book information makes it possible to determine the bid-ask spreads prevailing

at the time of the trade. Since the data do not identify the initiating dealer, the method of Lee

and Ready (1991) is used to sign the trades. Trades that are executed at price less than the

mid price are classi�ed as seller-initiated, and trades that are executed at price higher than the

mid price are classi�ed as buyer-initiated. For trades executed at the mid price exactly, the

tick rule is used . The signed trades are then aggregated into daily net order �ow. Net order

�ow is de�ned as the number of buyer-initiated trades minus the number of net seller-initiated

trades during a day. Order �ow is thus a measure of the net buying pressure in the market.

Order �ow are divided into three segments according to the remaining time to maturity of the

bonds included. Short term order �ow includes the order �ow in bonds with a remaining time

to maturity between 1 and 4 years, medium term order �ow includes bonds with a remaining

time to maturity between 4 and 7 years and long term order �ow includes trades in bonds

with a remaining time to maturity between 7 and 11 years. The bonds included in long term

order �ow are less liquid than the bonds included in the other two order �ow categories. This

is because new Norwegian Government bonds are issued every second year as 11 year bonds

with a small outstanding volume. The volume of the new bonds is gradually increasing in size

through subsequent auctions over the next 6 - 7 years. There are typically 5 -10 bond auctions

in di¤erent issues during a year, this means that it takes time before the bond is liquid and

reaches its maximum size. In the secondary market the bid-ask spread is larger and size of

trades in the electronic order book smaller than for the shorter maturity bonds.

Zero coupon yields and forward rates for Norwegian government bonds are kindly provided

by Nordea Markets. These yields are calculated from end-of-day yields of the benchmark

government bonds and bills using the Nelson-Siegel algorihtm. Zero-coupon bond yields with

maturities of 1, 2, 3, 4, 5 and 10 years are used to calculate daily and monthly yield changes,

excess returns, forward spreads and principal components of forward rates. Maturities of 1-5

years are included to correspond to the data included in Fama and Bliss (1987) and Cochrane

and Piazzesi (2005). The 10 year yield is included in the analysis to explore whether the long

end of the yield curve behaves in the same way as the short and medium term part of the curve.

The secondary market for Norwegian government bonds consists of an interdealer market

9

and a customer market. The interdealer market is limited to members of OSE that are ap-

proved for bond trading. Typical members will be banks and brokerage �rms. The majority of

interdealer trading activity involves primary dealers. The primary dealers have speci�c rights

and obligations, and will provide tradable bid and ask prices continuously in the electronic

trading system throughout the trading day. In the interdealer market trading takes place both

in the electronic and the over-the-counter market. In the customer market trades are made

in the over-the-counter market. Customers including commercial enterprises and institutional

investors have to execute a bond trade trough an authorized bond member of the OSE.

Trading in the secondary market for Norwegian government bonds mainly takes place be-

tween 9 am and 4 pm local time. Trades that are executed outside of these hours will be entered

into the trading system, Saxess, around 9 am. The interdealer trades constitute about 35 per-

cent and customer trades about 65 percent of the total market measured in number of trades.

The share of electronic trading and the average electronic trade size has gradually increased

since the inception of an electronic order book in 1999. Still, a majority of interdealer trades

and all customer trades are over-the-counter trades. These trades are reported into the OSE

trading system within 5 minutes and will then be visible to other traders. The only exception

is delayed publication trades which are also reported within 5 minutes, but not visible for the

other traders until after a delay. The period of delay and the conditions for delaying a trade

have changed during the sample period. I 1999 only trades over 250 million NOK were allowed

registered as delayed publication trades, and the delay was 1 hour only. In 2005 the size limit

was abandoned and the delay extended to the end of the trading day at 4 pm. This delay is

meant to allow market makers to unwind their inventory positions at minimal cost. By hid-

ing these trades from the other market participants, the users of delayed publication trades

become temporary monopolists on trade information and can update their beliefs before the

other traders. The system of delayed publication thus leads to a less transparent market and

information asymmetries occur.

The time for private information to become incorporated into prices depends on market

conditions. Market liquidity, which refers to the ease by which one can trade without large

price impacts, is one condition of importance. Brandt and Kavajecz (2004) �nd that the e¤ect

of order �ow on yields is strongest when liquidity is low. This implies that price discovery takes

less time in less liquid markets, on the other hand, there has to be a certain level of trading in

order to facilitate the price discovery process. Market transparency is another market condition

which is of importance for the price discovery process. Transparency refers to the ability to

observe the information in the trading process. This depends on the trading rules of the market,

e.g. the ability to observe individual trades or net order �ow only, to observe the limit order

book for the whole market or own order book only. The possiblilty to delay the publication

of a trade is another trading rule in�uencing the transparency of a market. In the Norwegian

soverign bond market dealers can hide their trades from other market participants by delaying

the publication of a trade until the end of the day. This type of trade will make a market less

transparent and implies that dealer will be an information monopolist for a limited time. As an

10

information monopolist she can update her beliefs before the other market participants make

transactions being better informed. Delaying the reporting of trade information leads to stale

prices and slows the process of price discovery in the market.

The descriptive statistics, including the AR(1), of the variables used in the predictive re-

gressions are presented in table 1a and 1b. Statistics for daily observations are displayed in

table 1a and statistics for non-overlapping monthly data, interest rates reported every 20th

day and accumulated 20 day order �ow, are displayed in table 1b. Due to an in�ation level

below target from the end of 2002, the Central Bank cut the key interest rate from 7 percent in

December 2002 to 1.75 percent in 2004. This led to a sharp decline in bond yields during the

sample period and the table show that mean yield changes are negative both on a daily and

monthly horizon. Excess returns are on average negative as well due to a downward sloping

yield curve for part of the sample period. It appears from the tables that the autocorrelation

of yield changes and excess returns follow the same pattern, with decreaseing autocorrelation

from 1 year to 5 year and slighly increasing up to the 10 year maturity. Of the daily order �ow

variables, medium term order �ow has the lowest standard error and the lowest autocorrelation

at less than 7 percent. Short and long term daily order �ow have a autocorrelation 15 and 11

percent respectively. This increases to levels above 30 percent at the monthly horizon. Forward

spreads and principal componenets are very persistent variables at both the daily and monthly

horizon.

Table 2a and 2b shows the decomposition of forward rates into principal components. Table

2a presents the six factors extracted from the six forward rates and shows that the �rst factor

explains 92.8 percent of total variance, whereas the second and third factors explain 6.6 and

0.5 percent respectively. Table 2b shows the loadings of the three factors on the di¤erent

forward rates with maturities from one to ten years. The �rst factor loads about equally

on all forward rates. This makes it comparable to the "level" factor described by Litterman

and Scheinkman (1991) whereas the second and third principal components corresponds to

the "slope" and "curvature" factors.2 Table 2c shows the correlation between the �rst three

principal components of forward rates and the �rst three principal components of spot rates.

The correlations between the �rst and second factors of forward rates and spot rates are high,

95 and 78 percent respectively, indicating that they capture about the same attributes of the

yield curve. The correlation between the third principal component of forward rates and the

third principal component of zero spot rates is much lower, only 18 percent, suggesting that

this forward factor represents an attribute speci�c to forward rates. Brandt and Kavajecz

(2004) and Valseth (2008) extracts principal components of yields from the Treasury market

and Norwegian government bond market respectively, and report that lagged yield factors do

not explain daily yield changes. Since only the third principal component di¤er markedly

between spot yields and forward rates, any predictive power in forward rates may be related to

2Litterman and Scheinkman (1991) determines the common factors that have a¤ected the Treasury returnsand �nd that the variation in returns on all Treasury �xed income securities can be explained by three factorsnamed level, steepness and curvature.

11

this factor.

1.3 Theoretical background and econometric framework

The classical expectations hypothesis, which constitute the foundation of interest rate pre-

dictability, states that bond yields are the expected average of future short interest rates,

implying that forward rates predict interest rate changes. Modi�ed versions of the expecta-

tions hypothesis include a constant risk premium, indicating that the forward rate minus a

constant predicts the future interest rate. Fama and Bliss (1987) and Campbell and Shiller

(1991) �nd that forward rates and yield spreads have little predictive power for future interest

rates and conclude that their �ndings are inconsistent with the expectations theory. Instead

they �nd that forward rates predict excess bond returns, a proxy for bond risk premia. Bond

risk premia, which are de�ned as expected returns on bonds over expected returns on short

riskless investments, compensate for interest rate risk and other types of risk related to gov-

ernment bonds. Cochrane and Piazzesi (2005) con�rm earlier �ndings by documenting that a

linear combination of forward rates have strong forecasting power for all bond excess returns.

These studies strongly indicate that bond yields contain both time-varying risk premia and

expected average short rates, and suggest that risk premia are predictable while future short

rates are largely unpredictable.

This paper investigates whether lagged order �ow have predictive power beyond that of

traditional term structure models. The models employed here are based on the simple regres-

sions of Fama and Bliss (1987). In the following paragraphs the relationship between prices

and yields on zero coupon bonds based on the expectations hypothesis are therefore brie�y

presented.

The relationship between the price and the yield of a zero-coupon bond with N years to

maturity is shown in equation (1).

P (N) = [1 + Y (N)]�N (1)

By taking logs on both sides of equation (1) and expressing yield as a function of price,

equation (2) displays the widely used relationship between the log yield and log price of a zero

coupon bond with N years to maturity.

y(N) = � 1Np(N) (2)

The one period log return of a bond with N years to maturity at the beginning of the period

and N-1 years to maturity at the end of the period is expressed in equation (3).

r(N)t+1 = p

(N�1)t+1 � p

(N)t (3)

12

The expectations hypothesis of the term structure of interest rates can be stated in three

di¤erent ways as in Cochrane (2001). He emphasizes that the hypothesis is "three equal state-

ments about the pattern of (zero) yields accross maturity". First, the expectations hypothesis

states that bond yields are expected values of average future short term (one-period) rates as

shown in equation (4).

y(N)t =

1

NEt(y

(1)t + y

(1)t+1 + y

(1)t+2:::y

(1)t+N�1) +Risk Premium (4)

Second, the hypothesis implies that forward rates equals expected future spot rates. Equa-

tion (5) says that the forward rate quoteded today for the one period interest rate running from

period N to period N+1 equals the expected one period spot rate at time N.

f(N!N+1)t = Et(y

(1)t+N) +Risk Premium (5)

Finally, the expectation hypothesis implies that the expected holding period return is the

same for bonds of all maturities, which is illustrated in equation (6).

Et(r(N)t+1) = y

(1)t +Risk Premium (6)

In its pure form there is no risk premium included i the expectations model, but the model

can be modi�ed to include a constant term premium. By rearranging equation (6) it appears

that the theoretical risk premiuim equals the expected excess return as shown in equation (7).

Et(exrNt+1) = Et(r

(N)t+1)� y

(1)t (7)

Two forecasting horizons are analyzed in this paper, one day and twenty days. The twenty

day horizon is also referred to as the monthly horizon in this paper. The daily and monthly

yield changes on six zero coupon bonds, bonds with 1, 2, 3, 4, 5,and 10 years maturity, are

forecasted in this study. Since the forecasting horizons are short relative to the maturity of

the bonds, the yield changes and excess returns are estimated under the assumption that the

remaining time to maturity of the bond is approximately the same at the beginning and at the

end of the forecasting period. It is thus assumed that N years � 1 day t N years and N

years� 1 month t N years: Yield changes and bond returns are thus calculated according to

equations (8) and (9) where t are measured as days in the daily analysis and as months in the

monthly analysis.

dy(N years)t+1 = y

(N years)t+1 � y

(N years)t (8)

13

r(N years)t+1 = p

(N years)t+1 � p

(N years)t (9)

From equation (6) it appears that the expected excess return is a measure of the risk

premium. Since expectations are unobservable, the actual excess return is used as a proxy for

the bond risk premium. Equation (10) displays how the excess bond returns are calculated.

The one month zero rate is used as a proxy for the one-period riskless return.

exr(N years)t+1 = r

(N years)t+1 � y

(1)t (10)

Several studies document that the bond risk premium is neither zero nor a constant, but

time-varying. Changes in yields can according to equation (4) therefore be a result of either

changes in expected future short rates or changes in the risk premium. Fama and Bliss (1987)

�nd that the forward spread tracks changes in risk premia. The forward spread, FSNt ; is de�ned

in equation (11) and is measured as the one period forward rate starting at time Nminus today�s

one period rate.

FSNt � f(N!N+1)t � y

(1)t (11)

Equation (12) shows that the forward spread equals the sum of the expected risk premium

and the expected yield change next period. Fama and Bliss (1987) also �nd that the forward

spread is a poor predictor of interest rate changes, but that the forecast power increases some-

what with the horizon. Since forward rates are poor forecasters of interest rates, they conclude

that yields are close to random walks. Equation (12) then suggests that the forward spread

should predict the risk premium. The main �nding of Fama and Bliss (1987) is that the forward

spread predicts realized excess return which is a a proxy for the theoretical risk premium.

FSNt = Et(exrNt+1) + Et(y

(N)t+1 � y

(N)t ) (12)

In order to control for the information imbedded in the yield curve when testing for order

�ow predictability, lagged forward rates are included as predictor variables in the regression

models. Therefore yield changes are �rst predicted using yield curve data only. Both the Fama-

Bliss maturity dependent forward spread de�ned in equation (11) and forward rates are used

as predictor variables. However, instead of using the Cochrane-Piazzesi linear combination of

forward rates, common factors are extracted from the six forward rates by the method of princi-

pal components. The three �rst principal components are included as explanatory variables in

the regression models. Table 2a and 2b shows the decomposition of forward rates into principal

components and reveals that the �rst three factors explain 99,9 percent of the variance. On

14

the daily forecast horizon the one equation model a to model e, where t denotes day, are used

to forecast daily yield changes, denoted dy(N years)t+1 , and daily risk premia denoted exr

(N years)t+1 .

Model a) presented in equation (13) use the Fama-Bliss maturity dependent forward spread as

the predictor variable.

dy(N years)t+1 = �0 + �1FSt + "t+1 (13)

Model b) presented in equation (14) include the �rst three principal components of forward

rates as predictive variables.

dy(N years)t+1 = �0 + �

1

2F1t + �

2

2F2t + �

3

2F3t + "t+1 (14)

Model a) and b) have forward rates only as explanatory variables. Model c) uses order �ow

only to predict. Order �ow is divided into three groups according to the remaining time to

maturity of the bonds traded. OF S refers to short term order �ow, which is the net order �ow

of interdealer trades in bonds with a remaining time to maturity of between 1 - 4 years. OFM

refers to medium term order �ow, which is the net order �ow of interdealer trades in bonds

with a remaining time to maturity of between 4 - 7 years. OFL refers to long term order �ow,

which is the net order �ow of interdealer trades in bonds with a remaining time to maturity of

between 7 - 11 years. Model c) presented in equation (15) includes only order �ow as predictor

variables. Model c) is also used in the out-of-sample analysis in order to isolate the predictive

power of order �ow.

dy(N years)t+1 = �0 + �S3OF

St + �M3 OF

Mt + �L3OF

Lt + "t+1 (15)

Model d) and e) are models a) and b) extended with lagged order �ow respectively, and

represented in equations (16) and (17).

dy(N years)t+1 = �0 + �1FSt + �S3OF

St + �M3 OF

Mt + �L3OF

Lt + "t+1 (16)

dy(N years)t+1 = �0 + �

1

2F1t + �

2

2F2t + �

3

2F3t + �S3OF

St + �M3 OF

Mt + �L3OF

Lt + "t+1 (17)

Another purpose of this paper is to investigate whether order �ow contains information

about bond risk premia. By testing the predictive power of order �ow on both yield changes and

excess bond returns, the empirical results can be analyzed in light of the expected theoretical

15

results. If order �ow contain information related to bond risk premia only, a prediction of an

increase in yields should according to equation (4) be parallelled by a prediction of an increase

in the risk premium. This requires that the sign of the order �ow coe¢ cients are equal in both

predictive regressions. If order �ow contain information related to both risk premia and future

short interest rates, a prediction of an increase in yields could imply a prediction of an increase

in the risk premium and an increase in future short rates, but it could also imply a fall in the

risk premium that is more than compensated by an expected increase in short rates. If the

latter is the case, the sign of the order �ow coe¢ cients could di¤er in the predictive regressions

of yield changes and of excess returns. All regression models are thus performed twice for each

forecast horizon: �rst on yield changes and then on excess returns.

At the monthly forecast horizon the following one equation models f - j, where t denotes

month, are used to forecast monthly yield changes, dy(N years)t+1 ;and monthly risk premia, exr(N years)

t+1 :

At the monthly horizon only one order �ow variable is included in the models. This variable is

the sum of short term and medium term order �ow and aggregeted over 20 day periods. The

long term order �ow is excluded as there are less transactions in long end of the bond market

than the short and medium term part. It is thus assumed that a majority of informed trades

takes place in the more liquid parts of the bond market.

Model f:

dy(N years)t+1 = �0 + �1FSt + "t+1 (18)

Model g:

dy(N years)t+1 = �0 + �

1

2F1t + �

2

2F2t + �

3

2F3t + "t+1 (19)

Model h:

dy(N years)t+1 = �0 + �3OF

S+Mt + "t+1 (20)

Model i:

dy(N years)t+1 = �0 + �1FSt + �3OF

S+Mt + "t+1 (21)

Model j:

dy(N years)t+1 = �0 + �

1

2F1t + �

2

2F2t + �

3

2F3t + �3OF

S+Mt + "t+1 (22)

In order to document the predictive power of bond market order �ow on yield changes

and excess returns, both in-sample and out-of-sample analyses are performed on both forecast

horizons.

16

1.4 In-sample results

The results of the in-sample predictions of yield changes and excess returns are presented in

this section. Order �ow has signi�cant predictive power for both yield changes and excess

returns. Forward rates, in line with the results of Fama and Bliss (1987) and Cochrane and

Piazzesi (2005), predict considerably more of excess returns than of yield changes. Both the

forward spread, which is the predictor variable in model a, and the three principal components

of forward rates, which are the predictor variables in model b, have predictive power for bond

risk premia on daily and monthly horizons. For daily predictions of yield changes, only the

one year forward spread have signi�cant predictive power for changes in the 1-year yield. For

monthly predictions of yield changes, the third principal component of forward rates appear to

have signi�cant predictive power.

The results of the in-sample predictions of yield changes and excess returns at a daily horizon

are presented in tables 3a and 3b respectively. According to table 3a forward rates appear to

have little power to predict daily yield changes except at the very short end. However, table

3b illustrates that forward rates have some forecasting power for daily excess returns. Table

3a ans 3b show that lagged order �ow has predictive power for both yield changes and excess

returns. Table 3a displays that medium term order �ow, which includes the trades in bonds

with a remaining time to maturity between 4 and 7 years, has signi�cant forecasting power for

yield changes of all maturities expect for the 10-year yield. This seems reasonable as this part

of the yield curve is more liquid than the long end, and informed dealers will thus prefer to take

positions in this segment. The explanatory power of models c, d and e are about the same,

indicating that only order �ow contains information about next day yield changes. The only

exception are the models of 1-year yield changes. Here the inclusion of the forward spread, in

addition to order �ow, increase the adjusted R2 from 2.9 percent in model c to 3.6 percent in

model d.

Table 3b shows that while order �ow has about the same explanatory power for yield changes

and excess returns, the explanatory power of forward rates is considerably larger, especially

for the daily excess return on 1-year bonds. Whereas model c explains 3 percent of daily

variation in the excess returns in 1-year bonds, model e explains 7.8 percent. Medium term

order �ow appears to have the strongest predictive ability on excess returns also. Comparing

the predictive power of models b, c and e it becomes clear that order �ow contains information

that is beyond the information in the current yield curve and that the predictive power of order

�ow is substantial.

Table 3c displays the role of delayed publication customer trades, or hidden customer order

�ow (HCOF), for the predictive power of interdealer order �ow and is based on equation ( ).

17

dyt+1= �0+�1F1t+�2F2t+�3F3t+�4HCOFSt +�5resOF

St +�6HCOF

Mt +�7resOF

Mt +�8HCOF

Lt +�9resOF

Lt +"t+1

(23)

The residual interdealer order �ow is extracted from equation ( ) for short, medium and

long term order �ow.

OF t= c+ HCOF t+�t ; �t= resOF t (24)

The table shows that delayed publication customer trades have signi�cant predictive power

for yield changes at the short end and long end of the yield curve, but not in the middle segment.

Table 4a and 4b displays the results of the in-sample predictions of monthly yield changes.

Forward rates have signi�canly higher predictive power for monthly excess returns than for

monthly yield changes. This is in line with the results from the daily predictions and the

empirical literature. However, unlike at the daily forecast horizon, it appears that the third

common factor of forward rates has a signi�cant impact on monthly yield changes. This princi-

pal component is relatively weakly correlated to the corresponing principal component of spot

rates, indicating that the third factor contains a particular feature related to forward rates that

contains information about future yield changes.

Only one order �ow component, which is the sum of the monthly short and medium term

order �ow, is included in the monthly regression models. This is because trading in these

segments are more liquid and therefore each contain a larger number of trades than the long

term order �ow. The combination of the short and medium term order �owmay thus be a better

measure of the monthly net buying pressure in the Norwegian government bond market than

including all maturities. Also, pricipal components analysis of Norwegian spot rates indicates

that 99 percent of the variation in yields is explained by the "level" factor. This implies that

yields normally change in parallell shifts and that the predictive power of order �ow will apply

to all yields. Table 4a show that order �ow explains a substantial part of monthly yield changes,

with signi�cant coe¢ cients at the 5 percent level or better and with an adjusted R2 of model

c�of between 8-17 percent. Thus the in-sample results are in line with Evans and Lyons (2005)

who �nd that the predictive power of lagged order �ow increases over longer horizons.

Table 4a furher conveys that both the third forward factor and order �ow have signi�cant

forecasting power. An increase in order �ow and a fall in the third principal component of

forward rates of one standard deviation each will predict a fall in the 3-year yield of 64 basis

points, or about 2/3 of a percetage point, the following month. The explanatory power of

model e�is greater than that of the other models and indicates that forward rates and order

�ow contain independent information relevant for future yield changes.

18

Table 4b conveys that lagged order �ow explains slightly more of monthly yield changes

than of monthly excess returns. The explanatory power of model c�varies from 8 - 13 percent.

However, since forward rates have higher forecasting power for excess returns than for yield

changes, model e�, including both forward rates and order �ow, has an adjusted R2 between

9 and 49 percent for monthly excess returns. An increase in the third principal component of

forward rates of one standard deviation predicts a fall in the monthly excess return on a 3-year

bond with 1.16 percent the next month, whereas a fall in monthly order �ow of one standard

deviation will decrease the monthly return on the 3-year bond with 0.70 percent.

The results reveal that the order �ow coe¢ cients have di¤erent signs for yield changes

and excess returns. This follows from the one-to-one relationship between yields and prices

on bonds shown in equation (1). According to table 3a and 3b a one standard deviation

increase in medium term order �ow will reduce the 5 year yield by 11 basis points tomorrow

and increase excess returns on the 5 year bond with 50 basis points. Net purchases of bonds

predict an increase in bond risk premia. This increase in the risk premium can be interpreted

as the systematic risk induced by private information discussed in Easley and O�Hara (2004).

Private information in order �ow increases the risk to uninformed investors holding the bond

as informed investors have the knowledge to shift their portfolio weigths to adjust to the new

private information. Excess returns will increase as an increase in bond prices induces a capital

gain on holding a bond.

1.5 Out-of-sample results

Out-of-sample forecasts from the three models b, c and e are based on recursive estimation

over the sample period September 1999 �September 2005. The daily forecasts of future yield

changes and excess returns are made for the period September 4, 2000 �September 30, 2005 and

the monthly data for the period August 2001 - September 2005. In model b with forward rates

as explanatory variables, only the third principal component is included in the monthly out-of-

sample predictions. This modi�cation is due to the better forecasting ability of this "forward-

speci�c" factor in the monthly in-sample results. To test the out-of-sample predictability of

these models, the mean-squared forecasting error (MSE) of the recursive forecasts from each

model is calculated and subsequently they are compared to the MSE of the random walk. The

random walk is a naïve forecasting method based on simple assumptions. One version of the

random walk hypothesis implies that today�s yield is the best prediction for next period�s yield

and thus gives the prediction of no change. This implies forecasts of daily yield change equal to

zero, weekly yield change equal to zero or monthly yield change equal to zero, depending on the

forecast horizon. An alternative version of the random walk is based on today�s historic average

of the yield change as the best predictor for the yield change in next period. This model has

the advantage of providing a daily updated forecast based on the latest information available

and a forecast that depends on the horizon, and is presented in equation (25). Intuitively

19

it appears more reasonable to expect no yield change from today until tomorrow than from

today until next month. Also, Goyal and Welch (2008) �nd some evidence that variables that

can predict in-sample cannot predict out-of-sample better than a constant, implying that the

historic average is a suitable benchmark.

dy(N years)t+1 = c+ �t+1 (25)

The following analysis is based on the random walk model de�ned in equation (25). In

order to test the out of sample predictive power of forward rates and order �ow, the forecasts

from the three models are compared to the forecasts from the random walk model. First the

ratio of the MSE of model b, c and e over the MSE of the random walk model is calculated. A

ratio less than one indicates that the alternative model beats the random walk. Then equality

of the mean squared forecasting error (MSE) of each model relative to the MSE of the random

walk is tested by using the McCracken (2007) MSE-F test. McCracken (2007) has developed a

test statistic which tests the null hypothesis that the constant yield change model has a MSE

that is less than, or equal to that of the time varying yield change model. The alternative

hypothesis is that the time-varying model has a lower MSE. The test statistic is presented in

equation (26)

MSE � F = (T � h+ 1) � (MSER �MSEUMSEU

) (26)

Where MSER is the mean squared forecast error of the random walk and MSEU is the mean

squared forecast error of the alternative model being tested. Equation (X2) de�nes the test

statistic as the the ratio of the di¤erence in the MSE of the model being evaluated and the

MSE of the random walk over the MSE of the model times the number of observations in the

sample, T; minus the horizon, h, plus one.

The results of the out-of-sample predictions for daily yield changes and excess returns are

reported in table 5a and 5b. The MSE-ratios and MSE-F tests for the three models predicting

daily yield changes are reported in table 5a. It appears that the order �ow model, model c,

signi�cantly outperforms the random walk for all maturities. All the MSE-f test statistics are

signi�cant at the 1 percent level or better. Model b, including the �rst three common factors

of the forward rates, do not perform well out-of-sample and is outperformed by the random

walk. Table 5b con�rms that lagged order �ow models signi�cantly outperforms the random

walk in predicting next day bond excess returns. For 1 year bonds the forward rate model also

outperforms the random walk signi�cantly. For longer bonds however, forward rates appear to

have no out-of-sample predictive power.

Figure 1 - 3 show the performance of the three models versus the random walk model

(RW). Figure 1a displays the predictive power of order �ow on daily changes in the 3 year yield

versus that of the random walk. The positive slope indicates that the accumulated MSE of the

20

order �ow model is smaller than the accumulated MSE of the random walk and shows that

the out-of-sample forecasts based on daily order �ow clearly outperforms the random walk over

the period September 4, 2000 to September 30, 2005. Figure 1b shows how well the principal

components of forward rates do in predicting the 3 year yield changes compared to the random

walk. The negative slope indicates that the forward rate model performs worse and that the

accumulated MSE of the model is greater than the accumulated MSE of the random walk.

Finally, �gure 1c demonstrate that a combined model including both order �ow and the three

forward rate factor does not perform well because of the poor predictive power of forward rates.

Table 6a and 6b presents the results of monthly predictions of yield changes and excess

returns based on non-overlapping data. Lagged short and medium term order �ow have sig-

ni�cant predictive power over the random walk with a MSE-ratio of 0.84 for the one year

monthly yield change. When including only the third principal component of forward rates,

the predictions of model b also signi�cantly outperforms the random walk for maturities of 2

-5 years. The combined model, including both order �ow and the third forward rate factor,

thus strongly outperforms the random walk for 2,3,4 and 5 year yield changes. Table 6b reveals

that order �ow can predict monthly excess returns out-of-sample too, but that the forecasts are

less signi�cant than for monthly yield changes. This suggests that monthly order �ow contain

information on future bond risk premia as well as on future short rates. As a robustness check

the monthly predictions are also performed on 20 day overlapping data, and the results are in

line with the results of the non-overlapping data, but not reported here.

Figure 4 - 6 illustrate the performance of the three models versus the random walk model

on a monthly horizon over the period August 2001 �September 2005. Figure 2a displays the

di¤erence in the accumulated MSE between the random walk and the order �ow model for

monthly 3 year changes. The positive slope indicates that the accumulated MSE of the order

�ow model is smaller than the accumulated MSE of the random walk. Figure 2b shows that

the third forward rate common factor has predictive power on the monthly horizon, indicating

that this factor contain information about expected yield changes. That a combined model

including both the sum of short and medium term order �ow and the third forward rate factor

is the best model is illustrated in �gure 2c. The slope of the di¤erence in accumulated MSE of

this model versus the MSE of the random walk is clearly steeper than for the other two models

separately, indicating a stronger outperformance of the random walk model.

The out-of-sample results presented in this section document that interdealer order �ow can

predict yield changes and excess bond returns on both daily and monthly horizons. Out-of-

sample tests con�rm that order �ow have signi�cant predictive power. The results from the

in-sample analysis indicate that the predictive power of order �ow is unrelated to risk premia.

The results in paper document that the predictive power of order �ow is roboust to possible

bias due to overlapping observations used in previous studies. Order �ow aggregated over

monthly horizons is a highly persistent variable when employing overlapping observations. Daily

21

order �ow aggregated over longer horizons may thus lead to biased coe¢ cients and R2 in

long horizon predictions. This problem is avoided here by including non-overlapping data

only. Beacuse of the relatively small sample of non-overlapping monthly data, the results are

controlled by performing monthly predictions based on overlapping data. The results of the

two methods give similar results indicating that there is no small sample bias.

22

1.6 Dealer heterogeneity

Since the information content in order �ow is due to the existence of informed traders, it

is interesting to investigate whether there are information asymmetries between the di¤erent

dealers in the market. An important contribution of this paper is thus to document that dealers

are heterogeneous with respect to the predictability of their order �ow. If dealers are di¤erently

informed, then the order �ow of well-informed dealers should be better predictors of future yield

changes/excess returns than the order �ow of less informed dealers. The information held by

di¤erent dealersmay vary because they have di¤erent sources information. The two most likely

sources of information are their customer trades and their own e¤ort and skill in interpreting

macroeconomic news and other information related to bond market developments. If the source

of information is customer trades, dealers may be di¤erently informed due to di¤erences in the

size and composition of their customer base. If dealer skill and e¤ort is the source of information

in their interdealer order �ow, di¤erences in size and sophistication of analytical team support,

the extent of proprietary trading activity and the skill of the individual traders may explain

di¤erences in the information content of order �ow. A unique data set including the identity of

the buying and the selling dealer facilitates the calculation of individual dealer order �ow and

thus the comparison of the information content in the order �ow of various dealers.

Seven dealers have been present in the Norwegian government bond market throughout

the whole sample period. By dealers we refer to banks and brokerage houses and their bond

traders. Dealers who have merged with others, dealers who strongly reduced their activity and

dealers who only sporadically traded in the market during the sample period are not included

in this analysis. The seven dealers are grouped according to three criteria; size, activity in

the interdealer market and the share of trades that are published with a lag. Size is de�ned

as a dealer�s total gross volume of trades measured in NOK in the customer market and the

interdealer market combined. The gross volume of interdealer trades only includes trades

initiated by the dealer. To the extent a dealer uses limit orders instead of market orders when

initiating a trade, her size in the interdealer market may be underestimated. However, the

use of automated trading was relatively small during the sample period and it is assumed that

the use of limit orders do not seriously a¤ect the chosen size measure. On average interdealer

trades constitutes about 35 percent of the total trades in the two markets.

Dealers who have a larger market share in the interdealer market than in the customer

market may be more active proprietary traders or have more competitive prices than those

with a smaller interdealer share. Activity in the interdealer market is here measured as the

ratio of a dealers�market share in the interdealer market over the dealer�s market share in the

customer market. A ratio above one indicates that the dealer has a lager share in the interdealer

market than in the customer market, indicating an active dealer. A dealer may be active in

the interdealer market in order to earn a pro�t on proprietary trading, infer information about

future yield changes from other dealers or in order to o­ oad risk incurred from trades initiated

23

by others. If a dealer is a competitive market maker and others trade on his prices and have

limited ability to take overnight risk, a larger proportion of interdealer trades will be undertaken

to balance dealer inventory. If the number of these uninformative trades are signi�cant they

will "pollute" the interdealer order �ow, and the dealer will appear to be less informed than

passive dealers who may not have this risk "constraint". A ratio below one indicates that the

dealer has limited activity in the interdealer market and a relatively large or active customer

base. It also indicates that the dealer is not actively seeking information by trading with other

dealers.

The trading system for Norwegian government bonds opens up for the possibility to delay

the publication of a trade until the end of the day. This means that a transaction executed and

entered into the trading system in the morning will not be visible to other participants in the

electronic trading system/order book until 4 pm. A dealer will most likely use this possiblity

when he believes a trade contains information that he wants to hide from the other market

participants. By hiding it he may be able to execute many pro�table interdealer trades before

it becomes known by other dealers. He may choose to enter both customer trades and interdealer

trades with delayed publication. During the period September 1999 - September 2005 more

than 70 percent of the trades with delayed publication were customer trades. Customer trades

may contain information about speculative or hedging positions that may be relevant for future

risk premia. The use of delayed publication trades varies among dealers, and one possible reason

for this is that the dealers will enter more delayed publication trades the more informed trades

they receive, indicating that the order �ow of these dealers will be more informed and have

better predictive power than that of other dealers.

Table 11 shows an overview of the dealer groups according to the three criteria. For each

criterium the dealers are divided into two groups. The table displays the number of dealers in

each group, the group members�total market share in Norwegian kroner, the group members

relative activity in the interdealer market and the group members�share of own trades entered

with delayed publication. The three dealers with a ratio of interdealer market share to customer

market share above 1 constitutes group 1 active dealers and the four dealers with a ratio of

less than 1 are included in group 2 passive dealers. In the size category the four dealers with

a market share considerably lower than 20 percent constitute group 3 small dealers and the

three dealers with a market share of 20 percent and more constitute group 4 large dealers. The

third criterium, the share of delayed or temporarily "hidden" trades, divides the dealers into

group 5 small share, which consists of the �ve dealers with a share of delayed publication trades

less than the average, and group 6 large share, which consists of two dealers which utilizes the

possiblity of delaying the revelation of their trades to other dealers in about 14 percent of their

total trades.

Tables 12, 14 and 16 display the in-sample predictive power of the order �ow of the di¤erent

groups of dealers. The comparative analysis is performed on monthly data and for yield changes

only, in order to limit the number of tables. The order �ow of each group is referred to as OF1

24

for group 1 and so on up to OF6 for group 6 and consists of the sum of the short and medium

term order �ow of the dealers in each group. The long term order �ow is not included because

there is less liquidity in long bonds with a maturity between 7 -11 years than in the shorter

maturities, thus dealers with information will most likely do their transactions in one of the

two shorter maturity groups.

From table 12 it appears that the order �ow of passive dealers, have stronger in-sample

predictive power for yield changes next month than active dealers. Table 13 con�rms the

�ndings in table 12. The order �ow of passive dealers have signi�cantly stronger out-of-sample

predictive power than that of active dealers, except at the short end of the yield curve. For

predictions of changes in the one year yield it is not possible to conclude whether the predictive

power of the order �ow of group 1 is signi�cantly di¤erent from the predictive power of the

order �ow of group 2. The table further shows that whereas the order �ow of passive dealers

gives signi�cantly better forecasts than the random walk model for all maturities, the order �ow

of active dealers, do not result in better forecasts than the random walk for maturities of three

years and up. Active dealers may have lower risk limits and thus their interdealer order �ow,

which can be due both to customer trades and market making, may contain more "noise" in

the form of uninformative transactions than the order �ow of less active dealers. There are thus

two possible explanations for the lower explanatory power of the order �ow of active dealers.

It could be that this "noise" disguises the possible extra information they receive from being

active in the interdealer market or it could be that there is little information to be extracted

from extra activity the interdealer market because the predictive power of interdealer order

�ow is mainly due to information in customer trades.

Table 14 shows that the in-sample predictive power of large dealers�order �ow is signi�cantly

larger than the order �ow of small dealers by explaining a larger part of the variation in yield

changes. Table 16 con�rms that out-of-sample forecasts based on the order �ow of large dealers

only signi�cantly outperforms the forecasts based on small dealer order �ow for all maturities

from 1 to ten years. Whereas predictions based on large dealers always outperforms the random

walk, predictions based on small dealer order �ow only outperforms forecasts based on the

random walk model for yield changes in one year and two year zero coupon bonds. These

�ndings suggest that large dealers are better informed than small dealers, which can be due

to better informed customers or to better skill due to more resources put into bond market

analysis.

The corresponding out-of-sample results are shown in tables 13, 15 and 17.

In table 16 it appears that the order �ow of dealers who often use to delay the publication

of their trades have higher explanatory power than for any of the other groups. The lagged

order �ow of dealers with a large share of delayed publication trades explain nearly 18 percent

of the variation in monthly changes in the two year yield. A one standard deviation increase

in group 6 order �ow will lead to a reduction in the two year yield of 71 basis points or about34percentage point. From table 17 it appears that the predictive model for two years yield

25

changes including the order �ow of dealers with a large share of delayed pubication trades

is superior to the random walk model. The accumulated MSE from the order �ow model is

only 82 percent of (18 percent less than) the accumulated MSE from the random walk model.

Table 17 shows that the predictive power of the dealers that most often "hide" their trades is

signi�cantly higher than that of the others for all maturities. This could be due to a relative

higher share of informed customers than others indicating that the main source of information

in interdealer order �ow is customer transactions. This indicates that some dealers are able

to identify their informed customers and choose to hide these trades from others in order to

bene�t from the private information. This may be because some dealers have more informed

customers than others or because some dealers have better skills in detecting and utilizing

private information from their customers.

Figure (8) illustrates the increase in predictive power over the random walk by adding the

order �ow of dealers with a low share of delayed publication trades. It appears that this model

performs better than the random walk. Figure (9) shows the increase in the in predictive

power over the random walk by adding the order �ow of dealers with a large share of delayed

publication trades. It appears to be increasing over the whole period and clearly outperforms

the random walk. Figure (10) depict the increase in predictive power by adding the order

�ow of group 6, consisting of dealers with a high share of delayed publication trades, to the

model used in Figure (8), indicating that the predictive power of the order �ow of this group

is signi�cantly higher than the predictive power of the order �ow of the other dealer group.

The �gure clearly indicates that the dealers who frequently uses the possiblilty to delay the

publication of a trade are better informed than dealers who use this possibility less often. This

indicates that dealers choose to enter trades with delayed publication when they consider the

trade to be informative.

The results of the individual dealer analysis suggest that dealers are heterogeneous because

they have di¤erent sources and amounts of information. This is most likely due to di¤erent types

of customers, but may also be a result of di¤erences in dealer skill and e¤ort. It appears that

large dealers have most informed customers. Some dealers are able to identify their informed

customers and therefore seek to hide these customer trades and their own informed interdealer

trades in order to prevent the information to quickly spread to other dealers. Varying dealer

strategies with respect to the e¤ort spent to protect information in informed trades may also

explain the dealer heterogeneity found in this study.

1.7 Conclusion

This paper tests the in-sample and out-of-sample predictive power of lagged order �ow in �xed

income markets while controlling for traditional term structure variables. The models apply

both forward rates and order �ow as predictor variables and compare the predictions from

26

these models with the random walk. The results document that lagged order �ow can predict

changes in bond yields, and that the information contained in order �ow is additional to the

information already imbedded in the yield curve. Forward rates appear to have no out-of-

sample predictive power on daily yield changes, but the third principal component of forward

rates, which appear to be unique to forward rates, has predictive power on the monthly horizon.

The predictive power of order �ow arises because the price formation process, in which asset

prices adjust to full information prices, do not happen instantaneously, but evolves in markets

over time. Interdealer order �ow re�ects private information that is not yet incorporated into

the yield curve, including delayed publication trades that are temporarily hidden from other

market participants.

A novel �nding of this study is that dealers are heterogeneous. Di¤erences in the predictive

power of the order �ow of various groups of dealers appear to be due to di¤erences in the

share of informed customers and varying dealer strategies. The results suggest that dealers

who actively hide their trades by delaying the publication of the trades until the end of the

day are better informed than other dealers. Since 70 percent of the trades that are entered

into the bond trading system with delayed publication are customer trades, this implies that

an important source of information in interdealer trades is informed customer trades. Dealers

who have informed customers may identify informed customer trades and then choose to hide

this information from other dealers by using the trade �ag delayed publication. However, the

results of the indiviual dealer analysis may also open up for skill and e¤ort as a source of

information in interdealer order �ow. Dealers actively hiding their trades may exert more e¤ort

than others in analysing customer order �ow and market conditions before deciding how to

enter their trades into the electronic trading system. Dealers with a high share of delayed

publication trades appear to have stronger predictive power than other dealers. This may be

due to di¤erent dealer strategies on how to bene�t from the private information contained in

customer trades.

By investigating the predictive power of order �ow on both yield changes and excess bond

returns the paper also attempts to determine the type of information contained in order �ow.

Forward rates appear to have better predictive power for bond excess returns than for yield

changes. Order �ow has similar predictive power for yield changes and bond excess returns.

Order �ow may thus contain information about future short rates as well as risk premia, but

the models used in this paper cannot separate the two sources of information.

This paper also documents that the predictive power of order �ow is roboust to possible

bias due to overlapping observations used in previous studies.

27

References

[1] Andersen, T., Bollerslev, T., Diebold, F. and Vega, C. (2007) Real-time price discovery in

global stock, bond and foreign exchange markets, Journal of International Economics 73

(2007), 251-277

[2] Boudoukh, Richardson and Whitelaw (2008) The Myth of Long Horizon Predictability,

Review of Financial Studies vol 21 number 4 July 2008 (1577-1605)

[3] Brandt, M.and K. Kavajecz (2004) Price Discovery in the U.S. Treasury Market: The

impact of Order�ow and Liquidity on the Yield curve, Journal of Finance 59, 2623-2654

[4] Campbell, J. and Shiller, R. (1991) Yield Spreads and Interest Rate Movements: A bird�s

eye view, Review of Economic Studies 58, 495-514

[5] Clark, T.E. and M.W. McCracken (2005) Evaluating direct multistep forecasts, Econo-

metric Reviews 24:369-404

[6] Cooper, I and Priestley, R (2009) Time-Varying Risk Premia and the Output Gap, Review

of Financial Studies 22, 2801-2833

[7] Cochrane, J. (2008) The dog that did not bark, Review of Financial Studies 21, 4, July

2008 1533-1575

[8] Cochrane, J. (2001) Asset pricing , Princeton University Press

[9] Cochrane, J. and Piazzesi, M. (2005) Bond Risk Premia, American Economic Review 95,

138-160

[10] Deaves, Richard (1998) Term premium determinants, return enhancement and interest rate

predictability, Journal of Business Finance & Accounting, 25(3) & (4) April/May 1998

[11] Easley, D. and O�Hara M. (2004) Information and Cost of Capital, Journal of Finance 59,

1553-1583

[12] Engsted, T. and Tanggaard, C. (1995) The Predictive Power of Yield Spreads for Fu-

ture Interest Rates: Evidence from the Danish Term Structure, Scandinavian Journal of

Economics 97(1), 145-159, 1995

[13] Engsted, T. and Tanggaard, C. (2001) The Danish stock and bond markets: comovements,

return predictability and variance decomposition. Journal of Empirical Finance 8, 243-271

[14] Evans and Lyons (2005) Meese-Rogo¤ Redux: Micro-based exchange rate forecasting ,

American Economic Review 2005, 95, 405-414

[15] Evans and Lyons (2008) Forecasting Exchange Rates and Fundamentals with Order Flow.

Working paper

28

[16] Evans and Lyons (2006) Understanding order �ow. International Journal of Finance and

Economics. 11,3-23

[17] Fama, Eugene and Bliss, Robert (1987) The information in long-maturity forward rates.

American Economic Review. 77, 680-692.

[18] Goyal, A and Welch, I (2008) A Comprehensive Look at The Empirical Performance of

Equity Premium Prediction, Review of Financial Studies 21, 4, July 2008 1455-1508

[19] Green, T. C., 2004, Economic news and the impact of trading on bond prices, Journal of

Finance 59, 1201-1233.

[20] Ilmanen, A. (1995) Time-varying expected returns in international bond markets, Journal

of Finance 50, 481-506.

[21] Kessler, S., and Scherer, B. (2009) Varying risk premia in international bond markets.

Journal of Banking anf Finance 33, 1361-1375

[22] Litterman, R., and Scheinkman, J., 1991, Common factors a¤ecting bond returns, Journal

of Fixed Income.1, 54-61.

[23] Ludvigson, S. and Ng, S. (2008) Macro Factors in Bond Risk Premia, Forthcoming in

Review of Financial Studies

[24] O�Hara, M. (1995) Market Microstructure Theory, Blackwell Publishing

[25] O�Hara, M. (2003) Presidental Address: Liquidity and Price Discovery, Journal of Finance

58, 1335-1354

[26] Pasquariello, P. and Vega, C. (2007) Informed and Strategic Order Flow in the Bond

Markets. Forthcoming in Review of Financial Studies

[27] Valseth, S (2008) An empirical study of the information content in order �ow in interdealer

and customer markets. Unpublished paper

29

Table 1aDescriptive statistics of daily data

Series obs mean std.err minimum maximum AR(1)

dm1 1504 -0.003 0.070 -0.58 0.56 -0.152

dy1 1504 -0.002 0.044 -0.37 0.25 0.156

dy2 1504 -0.002 0.052 -0.46 0.31 0.139

dy3 1504 -0.002 0.053 -0.49 0.29 0.107

dy4 1504 -0.002 0.052 -0.46 0.29 0.087

dy5 1504 -0.002 0.050 -0.42 0.29 0.082

dy10 1504 -0.002 0.046 -0.24 0.21 0.095

exr1 1504 -0.0002 0.0004 �0.003 0.003 0.156

exr2 1504 -0.0002 0.0010 -0.006 0.009 0.138

exr3 1504 -0.0002 0.0015 -0.009 0.014 0.107

exr4 1504 -0.0002 0.0020 -0.012 0.017 0.087

exr5 1504 -0.0002 0.0024 -0.014 0.020 0.082

exr10 1504 -0.0001 0.0043 -0.020 0.023 0.095

OFs 1504 -0.75 3.78 -27 22 0.154

OFm 1504 -0.17 3.69 -26 28 0.066

OFl 1504 -1.30 4.46 -34 20 0.113

fwd pc1 1504 -0.00086 2.36 -5.16 3.76 0.997

fwd pc2 1504 0.00024 0.63 -1.44 1.37 0.991

fwd pc3 1504 0.00002 0.17 -0.91 0.95 0.892

fwd spread1 1504 0.030 0.75 -2.08 1.52 0.990

fwd spread2 1504 0.252 1.18 -1.90 2.62 0.997

fwd spread3 1504 0.480 1.48 -1.80 3.43 0.998

fwd spread4 1504 0.640 1.66 -1.84 3.91 0.999

fwd spread5 1504 0.740 1.76 -1.86 4.21 0.999

fwd spread10 1504 0.875 1.85 -1.94 4.45 0.998

30

Table 1bDescriptive statistics monthly data

Series obs mean std.err minimum maximum AR(1)

dm1 74 -0.055 0.29 -1.13 0.53 0.157

dy1 74 -0.046 0.27 -0.98 0.47 0.382

dy2 74 -0.042 0.29 -0.80 0.56 0.244

dy3 74 -0.040 0.29 -0.69 0.64 0.173

dy4 74 -0.039 0.27 -0.68 0.67 0.151

dy5 74 -0.038 0.26 -0.63 0.67 0.154

dy10 74 -0.037 0.22 -0.53 0.61 0.153

exr1 74 -0.0035 0.0032 �0.0098 0.0064 0.577

exr2 74 -0.0031 0.0058 -0.0149 0.0123 0.309

exr3 74 -0.0028 0.0084 -0.0200 0.0156 0.209

exr4 74 -0.0025 0.0105 -0.0274 0.0214 0.178

exr5 74 -0.0021 0.0124 -0.0338 0.0255 0.177

exr10 74 -0.0004 0.0213 -0.0601 0.0463 0.169

OFs 74 -14.93 28.45 -97 70 0.350

OFm 74 -3.38 21.06 -77 44 0.308

OFl 74 -25.36 28.02 -111 22 0.348

fwd pc1 74 -0.027 2.38 -4.93 3.76 0.932

fwd pc2 74 0.011 0.62 -1.19 1.24 0.826

fwd pc3 74 -0.008 0.20 -0.84 0.76 0.468

fwd spread1 74 0.041 0.73 -1.67 1.46 0.886

fwd spread2 74 0.256 1.15 -1.64 2.51 0.948

fwd spread3 74 0.473 1.45 -1.58 3.30 0.968

fwd spread4 74 0.625 1.62 -1.60 3.72 0.975

fwd spread5 74 0.719 1.71 -1.63 3.94 0.977

fwd spread10 74 0.847 1.79 -1.67 4.05 0.976

31

Table 2aPrincipal components analysis of forward rates

Principal Value Proportion Cumulative AR(1)

component proportion

fwd pc1 5.569 0.928 0.928 0.997

fwd pc2 0.398 0.066 0.994 0.991

fwd pc3 0.030 0.005 0.999 0.892

fwd pc4 0.004 0.001 1.000 0.893

fwd pc5 0.000 0.000 1.000 0.881

fwd pc6 0.000 0.000 1.000 0.876

32

Table 2bLoadings of principal components on forward rates

fwd pc1 fwd pc2 fwd pc3

f(1 year!1 year+1m)t 0.388 0.630 0.403

f(2 years!2 years+1m)t 0.410 0.391 -0.046

f(3 years!3 years+1m)t 0.422 0.074 -0.347

f(4 years!4 years+1m)t 0.420 -0.189 -0.390

f(5 years!5 years+1m)t 0.412 -0.355 -0.258

f(10 years!10 years+1m)t 0.396 -0.532 0.705

33

Table 2cCorrelations between principal components of forward rates and principal

components of spot ratesfwd pc1 fwd pc2 fwd pc3 spt pc1 spt pc1 spt pc1

fwd pc1 1.000

fwd pc2 0.000 1.000

fwd pc3 0.000 0.000 1.000

spt pc1 0.951 0.285 0.061 1.000

spt pc2 0.304 -0.780 -0.231 0.000 1.000

spt pc3 0.066 -0.542 0.178 0.000 0.000 1.000

34

Table 3aIn-sample predictions of daily yield changes

35

FSt F1t F2t F3t OF St OFMt OFLt Adj:R2

dy(1 y)t+1 a 0:58

(3:02)0:008

b 0:00(0:53)

0:00(1:61)

�0:01(�0:64)

0:001

c �0:09(�2:05)

�0:13(�3:86)

�0:05(�1:54)

0:029

d 0:54(2:79)

�0:08(�1:78)

�0:13(�3:94)

�0:06(�1:86)

0:036

e 0:00(0:05)

0:00(1:49)

�0:00(�0:28)

�0:09(�1:97)

�0:13(�3:79)

�0:05(�1:59)

0:028

dy(2 y)t+1 a 0:12

(0:96)0:000

b �0:00(�0:06)

0:00(0:66)

0:00(0:38)

0:000

c �0:09(�1:74)

�0:13(�2:98)

�0:05(�1:62)

0:020

d 0:11(0:94)

�0:09(�1:71)

�0:13(�2:93)

�0:06(�1:64)

0:019

e �0:00(�0:56)

0:00(0:60)

0:01(0:76)

�0:09(�1:76)

�0:13(�2:95)

�0:06(�1:73)

0:018

dy(3 y)t+1 a 0:02

(0:19)0:000

b �0:00(�0:23)

0:00(0:52)

0:01(1:20)

0:000

c �0:07(�1:57)

�0:12(�2:75)

�0:05(�1:32)

0:014

d 0:02(0:25)

�0:07(�1:31)

�0:12(�2:53)

�0:05(�1:56)

0:013

e �0:00(�0:70)

0:00(0:48)

0:01(1:62)

�0:07(�1:39)

�0:12(�2:53)

�0:05(�1:74)

0:014

dy(4 y)t+1 a �0:01

(�0:11)0:000

b �0:00(�0:28)

0:00(0:62)

0:01(1:61)

0:001

c �0:04(�0:87)

�0:12(�2:33)

�0:05(�1:77)

0:011

d �0:00(�0:02)

�0:04(�0:87)

�0:12(�2:33)

�0:05(�1:75)

0:011

e �0:00(�0:65)

0:00(0:61)

0:02(1:87)

�0:05(�0:97)

�0:12(�2:33)

�0:06(�1:98)

0:013

dy(5 y)t+1 a �0:02

(�0:28)0:000

b �0:00(�0:23)

0:00(0:80)

0:01(1:48)

0:000

c �0:02(�0:46)

�0:11(�2:20)

�0:06(�2:17)

0:010

d �0:01(�0:18)

�0:02(�0:46)

�0:11(�2:21)

�0:06(�2:14)

0:010

e �0:00(�0:57)

0:00(0:82)

0:01(1:75)

�0:02(�0:55)

�0:11(�2:20)

�0:07(�2:40)

0:012

dy(10 y)t+1 a �0:07

(�1:00)0:000

b 0:00(0:04)

0:00(1:68)

�0:02(1:90)

0:003

c 0:03(0:74)

�0:06(�1:54)

�0:11(�4:01)

0:013

d �0:19(�1:21)

0:02(0:67)

�0:06(�1:50)

�0:10(�4:15)

0:013

e �0:00(�0:25)

0:00(1:86)

�0:01(�1:85)

0:03(0:88)

�0:06(�1:45)

�0:11(�4:12)

0:01636

37

Table 3bIn-sample predictions of daily excess returns

38

FSt F1t F2t F3t OF St OFMt OFLt Adj:R2

exr(1 y)t+1 a 0:35

(1:51)0:003

b �0:00(�8:33)

�0:01(�4:32)

�0:00(�0:25)

0:053

c 0:10(2:19)

0:12(3:45)

0:07(2:14)

0:030

d 0:40(1:76)

0:10(2:27)

0:12(3:73)

0:05(1:68)

0:034

e �0:00(�7:92)

�0:01(�4:52)

�0:00(�0:64)

0:07(1:70)

0:12(3:56)

0:06(1:92)

0:078

exr(2 y)t+1 a 0:53

(2:24)0:003

b �0:003(�3:41)

�0:008(�1:83)

�0:001(�0:83)

0:007

c 0:18(1:80)

0:24(2:82)

0:12(1:92)

0:020

d 0:54(2:38)

0:18(1:86)

0:25(2:95)

0:11(1:61)

0:023

e �0:003(�3:13)

�0:007(�1:87)

�0:02(�1:19)

0:16(1:63)

0:24(2:84)

0:12(1:90)

0:026

exr(3 y)t+1 a 0:56

(2:05)0:002

b �0:003(�2:02)

�0:008(�1:30)

�0:04(�1:59)

0:004

c 0:19(1:35)

0:35(2:45)

0:16(1:79)

0:014

d 0:55(2:13)

0:19(1:36)

0:36(2:52)

0:15(1:58)

0:016

e �0:002(�1:75)

�0:008(�1:31)

�0:05(�1:88)

0:19(1:29)

0:35(2:47)

0:17(1:86)

0:018

exr(4 y)t+1 a 0:59

(1:84)0:002

b �0:003(�1:45)

�0:01(�1:23)

�0:06(�1:81)

0:003

c 0:16(0:89)

0:43(2:26)

0:22(1:95)

0:011

d 0:56(1:86)

0:16(0:88)

0:44(2:32)

0:21(1:79)

0:013

e �0:002(�1:20)

�0:01(�1:25)

�0:07(�2:08)

0:16(0:88)

0:43(2:28)

0:23(2:08)

0:015

dy(5 y)t+1 a 0:63

(1:71)0:001

b �0:003(�1:18)

�0:01(�1:30)

�0:07(�1:65)

0:002

c 0:10(0:48)

0:50(2:14)

0:31(2:32)

0:011

d 0:59(1:68)

0:10(0:47)

0:51(2:20)

0:30(2:18)

0:012

e �0:002(�0:94)

�0:01(�1:35)

�0:07(�1:92)

0:10(�0:63)

0:50(2:16)

0:33(2:48)

0:013

exr(10 y)t+1 a 1:17

(1:75)0:002

b �0:004(�0:77)

�0:03(1:96)

0:15(1:81)

0:004

c �0:24(�0:71)

0:59(1:51)

1:04(4:07)

0:013

d

e �0:003(�0:53)

�0:04(�2:15)

0:13(1:75)

�0:31(�0:92)

0:57(1:44)

1:03(4:15)

0:017

39

Table 3cIn-sample predictions of daily yield changes based on delayed publication

customer trades.

The table presents the results of regressing yield changes on day t+1 on customer order �ow

based on delayed publication trades only (HCOF) and the residual interdealer order �ow

(resOF) of short term, medium term and long term maturity at time t. The regressions also

include a constant and the three �rst forward rate factors at time t as presented in the

equation below, but the coe¢ cients are dropped from the table. Coe¢ cients in bold

aresigni�cant at the 10 percent level.

dyt+1 = �0 + �1F1t + �2F2t + �3F3t + �4HCOFSt + �5resOF

St + �6HCOF

Mt + �7resOF

Mt +

�8HCOFLt + �9resOF

Lt + "t+1

HCOF St resOF St HCOFMt resOFMt HCOFLt resOFLt Adj:R2

dy1yt+1 �0:002(�1:91)

�0:001(�2:55)

�0:000(�0:27)

�0:001(�3:86)

�0:001(�0:93)

�0:000(�1:43)

0:028

dy2yt+1 �0:002(�1:75)

�0:001(�2:23)

�0:001(�0:82)

�0:001(�3:01)

�0:001(�1:23)

�0:001(�1:40)

0:018

dy3yt+1 �0:001(�1:07)

�0:001(�1:69)

�0:001(�0:85)

�0:001(�2:71)

�0:001(�1:38)

�0:001(�1:26)

0:014

dy4yt+1 �0:001(�0:49)

�0:000(�1:16)

�0:001(�0:77)

�0:001(�2:56)

�0:001(�1:62)

�0:001(�1:36)

0:011

dy5yt+1 �0:000(�0:08)

�0:000(�0:70)

�0:001(�0:67)

�0:001(�2:44)

�0:001(�1:95)

�0:001(�1:61)

0:011

dy10yt+1 0:001(0:77)

0:000(0:73)

�0:000(�0:27)

�0:001(�1:57)

�0:002(�3:22)

�0:001(�2:90)

0:017

40

Table 4aIn-sample predictions of monthly yield changes (non-overlapping data)

41

FSt F1t F2t F3t OFt Adj:R2

dy(1 year)t+1 a 11:8

(2:30)0:091

b 0:00(0:05)

0:05(0:92)

0:17(2:16)

0:000

c �0:26(�3:95)

0:175

d 7:60(1:79)

�0:23(�3:06)

0:203

e �0:01(�0:84)

0:02(0:54)

0:25(2:25)

�0:29(�4:11)

0:197

dy(2 years)t+1 a 2:62

(0:81)0:000

b �0:00(�0:34)

0:02(0:41)

0:26(2:47)

0:000

c �0:25(�3:45)

0:133

d 1:21(0:45)

�0:25(�3:05)

0:123

e �0:02(�1:33)

�0:01(�0:12)

0:35(2:38)

�0:28(�3:65)

0:183

dy(3 years)t+1 a 0:64

(0:27)0:000

b �0:01(�0:46)

0:02(0:41)

0:29(2:57)

0:011

c �0:23(�2:82)

0:108

d 0:25(0:12)

�0:23(�2:81)

0:096

e �0:02(�1:31)

�0:00(�0:11)

0:37(2:21)

�0:27(�3:50)

0:166

dy(4 years)t+1 a �0:01

(�0:01)0:000

b �0:00(�0:44)

0:03(1:76)

0:28(2:51)

0:015

c �0:20(�2:64)

0:096

d �0:09(�0:05)

�0:20(�2:65)

0:084

e �0:01(�1:18)

0:00(0:07)

0:35(2:59)

�0:23(�2:99)

0:154

dy(5 years)t+1 a �0:35

(�0:17)0:000

b �0:01(�0:41)

0:03(0:71)

0:26(2:56)

0:013

c �0:19(�2:54)

0:091

d �0:31(�0:18)

�0:19(�2:54)

0:079

e �0:01(�1:05)

0:01(0:28)

0:32(2:70)

�0:22(�2:80)

0:142

dy(10 years)t+1 a �1:14

(�0:83)0:000

b �0:00(�0:26)

0:06(1:60)

0:07(0:90)

0:000

c �0:15(�2:21)

0:076

d �1:04(�0:67)

�0:15(�2:17)

0:070

e �0:01(�0:74)

0:04(1:31)

0:12(1:73)

�0:17(�2:17)

0:08142

Table 4bIn-sample predictions of monthly excess returns. Non-overlapping data

43

model FSt F1t F2t F3t OFt Adj:R2

xr(1 year)t+1 a 6:60

(0:87)0:011

b �0:07(�6:08)

�0:14(�3:40)

�0:29(�3:42)

0:411

c 0:27(3:41)

0:136

d 12:94(1:99)

0:34(4:28)

0:211

e �0:06(�7:27)

�0:13(�3:55)

�0:35(�2:82)

0:21(3:38)

0:492

xr(2 years)t+1 a 10:28

(1:60)0:029

b �0:06(�2:44)

�0:14(�1:43)

�0:63(�2:88)

0:109

c 0:50(3:56)

0:134

d 13:42(2:84)

0:55(3:87)

0:194

e �0:05(�2:30)

�0:10(�1:15)

�0:77(�2:53)

0:50(3:46)

0:240

xr(3 years)t+1 a 10:87

(1:61)0:023

b �0:05(�1:47)

�0:16(�1:11)

�0:96(�2:83)

0:062

c 0:66(3:23)

0:110

d 12:07(2:32)

0:68(3:29)

0:143

e �0:03(�0:99)

�0:09(�0:76)

�1:16(�2:54)

0:70(3:24)

0:185

xr(4 years)t+1 a 11:53

(1:54)0:018

b �0:05(�1:06)

�0:20(�1:11)

�1:21(�2:91)

0:049

c 0:79(3:06)

0:098

d 11:83(1:98)

0:80(3:03)

0:120

e �0:02(�0:49)

�0:12(�0:74)

�1:46(�2:63)

0:86(3:09)

0:164

xr(5 years)t+1 a 12:58

(1:51)0:017

b �0:05(�0:85)

�0:25(�1:19)

�1:36(�3:00)

0:041

c 0:91(2:97)

0:093

d 12:40(1:81)

0:91(2:91)

0:110

e �0:01(�0:28)

�0:16(�0:84)

�1:65(�2:75)

0:99(2:98)

0:150

xr(10 years)t+1 a 21:29

(1:65)0:018

b �0:04(�0:46)

�0:66(�1:95)

�0:82(�1:22)

0:005

c 1:47(2:66)

0:077

d 21:42(1:82)

1:45(2:56)

0:093

e 0:01(0:07)

�0:52(�1:68)

�1:25(�1:88)

1:50(2:46)

0:085

44

Table 5aOut-of-Sample predictions of daily yield changes

Coe¢ cients in bold indicates a signi�cance level of 5 percent or betterForecast sample Comparison MSEU=MSER Test statistic 95 % cr. val.

dy(1Y )t+1

4.Sep.00-30.Sep.05 Order �ow vs const 0.98 34.38 -0.14

" Forward fact. vs const 1.01 -9.12 -0.14

" OF+Fwd.fact vs const 0.98 24.06 -3.44

dy(2Y )t+1

4.Sep.00-30.Sep.05 Order �ow vs const 0.98 22.24 -0.14

" Forward fact. vs const 1.01 -13.46 -0.14

" OF+Fwd.fact vs const 0.99 9.06 -3.44

dy(3Y )t+1

4.Sep.00-30.Sep.05 Order �ow vs const 0.99 14.34 -0.14

" Forward fact. vs const 1.01 -12.37 -0.14

" OF+Fwd.fact vs const 1.00 2.83 -3.44

dy(4Y )t+1

4.Sep.00-30.Sep.05 Order �ow vs const 0.99 10.65 -0.14

" Forward fact. vs const 1.01 -11.7 -0.14

" OF+Fwd.fact vs const 1.00 -0.26 -3.44

dy(5Y )t+1

4.Sep.00-30.Sep.05 Order �ow vs const 0.99 9.70 -0.14

" Forward fact. vs const 1.01 -12.89 -0.14

" OF+Fwd.fact vs const 1.00 -2.51 -3.44

dy(10Y )t+1

4.Sep.00-30.Sep.05 Order �ow vs const 0.99 15.51 -0.14

" Forward fact. vs const 1.01 -14.35 -0.14

" OF+Fwd.fact vs const 1.00 0.60 -3.44

45

Table 5bOut-of-Sample predictions of daily excess returns

Forecast sample Comparison MSEU=MSER MSE-F stat 95 % cr.val.

exr(1Y )t+1

4.Sep.00-30.Sep.05 Order �ow vs const 0.97 36.57 -0.14

" Forward fact. vs const 0.95 65.06 -0.14

" OF+Fwd.fact vs const 0.93 97.60 -3.44

exr(2Y )t+1

4.Sep.00-30.Sep.05 Order �ow vs const 0.98 23.36 -0.14

" Forward fact. vs const 1.00 -1.33 -0.14

" OF+Fwd.fact vs const 0.98 20.38 -3.44

exr(3Y )t+1

4.Sep.00-30.Sep.05 Order �ow vs const 0.99 14.95 -0.14

" Forward fact. vs const 1.01 -7.15 -0.14

" OF+Fwd.fact vs const 0.99 7.61 -3.44

exr(4Y )t+1

4.Sep.00-30.Sep.05 Order �ow vs const 0.99 11.14 -0.14

" Forward fact. vs const 1.01 -8.31 -0.14

" OF+Fwd.fact vs const 1.00 3.02 -3.44

exr(5Y )t+1

4.Sep.00-30.Sep.05 Order �ow vs const 0.99 10.19 -0.14

" Forward fact. vs const 1.01 -10.10 -0.14

" OF+Fwd.fact vs const 1.00 0.27 -3.44

exr(10Y )t+1

4.Sep.00-30.Sep.05 Order �ow vs const 0.99 16.02 -0.14

" Forward fact. vs const 1.01 -15.53 -0.14

" OF+Fwd.fact vs const 1.00 -0.84 -3.44

46

Table 6aOut-of-Sample predictions of monthly yield changes

Forecast sample Comparison MSEU=MSER MSE-F stat 95 % asym. distr

dy(1Y )t+1

Aug. 01- Sep.05 Order �ow vs const 0.84 10.35 1.46

" Forward fact. vs const 1.00 0.23 1.46

" OF+Fwd.fact vs const 0.80 13.09 1.52

dy(2Y )t+1

Aug. 01- Sep.05 Order �ow vs const 0.87 8.26 1.46

" Forward fact. vs const 0.97 1.78 1.46

" OF+Fwd.fact vs const 0.80 13.32 1.52

dy(3Y )t+1

Aug. 01- Sep.05 Order �ow vs const 0.89 6.46 1.46

" Forward fact. vs const 0.96 2.49 1.46

" OF+Fwd.fact vs const 0.82 12.00 1.52

dy(4Y )t+1

Aug. 01- Sep.05 Order �ow vs const 0.90 5.58 1.46

" Forward fact. vs const 0.95 2.62 1.46

" OF+Fwd.fact vs const 0.83 10.94 1.52

dy(5Y )t+1

Aug. 01- Sep.05 Order �ow vs const 0.91 5.22 1.46

" Forward fact. vs const 0.96 2.40 1.46

" OF+Fwd.fact vs const 0.84 10.09 1.52

dy(10Y )t+1

Aug. 01- Sep.05 Order �ow vs const 0.93 4.01 1.46

" Forward fact. vs const 1.00 0.03 1.46

" OF+Fwd.fact vs const 0.91 5.00 1.52

47

Table 6bOut-of-Sample predictions of monthly excess bond returns

Forecast sample Comparison MSEU=MSER MSE-F stat 95 % asym. distr

exr(1Y )t+1

Aug. 01- Sep.05 Order �ow vs const 0.86 8.60 1.46

" Forward fact. vs const 0.97 1.54 1.46

" OF+Fwd.fact vs const 0.79 13.83 1.52

exr(2Y )t+1

Aug. 01- Sep.05 Order �ow vs const 0.86 8.86 1.46

" Forward fact. vs const 0.95 3.04 1.46

" OF+Fwd.fact vs const 0.76 16.81 1.52

exr(3Y )t+1

Aug. 01- Sep.05 Order �ow vs const 0.88 7.02 1.46

" Forward fact. vs const 0.94 3.50 1.46

" OF+Fwd.fact vs const 0.78 14.69 1.52

exr(4Y )t+1

Aug. 01- Sep.05 Order �ow vs const 0.90 6.01 1.46

" Forward fact. vs const 0.94 3.42 1.46

" OF+Fwd.fact vs const 0.80 12.97 1.52

exr(5Y )t+1

Aug. 01- Sep.05 Order �ow vs const 0.91 5.55 1.46

" Forward fact. vs const 0.95 3.05 1.46

" OF+Fwd.fact vs const 0.82 11.68 1.52

exr(10Y )t+1

Aug. 01- Sep.05 Order �ow vs const 0.93 4.13 1.46

" Forward fact. vs const 1.00 0.24 1.46

" OF+Fwd.fact vs const 0.91 5.56 1.52

48

Table 7Groups of dealers according to di¤erent criteria

Group Criteria Group Number Total Relative Share of

category of market interdealer trades w/

dealers share share delayed publ.

1 Relative activity Active 3 32.6 % 1.52 10.6 %

2 in interdealer market Passive 4 67.4 % 0.79 9.4 %

3 Size measured as Small dealer 4 32.8 % 1.19 6.2 %

4 total market share Large dealer 3 67.2 % 0.91 11.9 %

5 Share of trades with Small share 5 55. 5% 0.90 7.0%

6 delayed publication Large share 2 44.5 % 1.10 14.1 %

49

Table 8In-sample predictions of monthly yield changes based on active vs. passive

dealers

coe¤. in bold are signi�cant at the 5 percent level or better

Maturity model FSt F1t F2t F3t OF1t OF2t Adj:R2

dy(1 year)t+1 a 8:18

(1:82)�0:21(�1:30)

�0:26(�1:77)

0:175

b �0:00(�0:34)

0:04(0:79)

0:24(2:41)

�0:39(�2:60)

0:110

c �0:00(�0:19)

0:05(1:13)

0:23(2:55)

�0:47(�3:38)

0:106

d �0:00(�0:47)

0:04(0:95)

0:27(2:58)

�0:29(�1:65)

�0:35(�2:41)

0:162

dy(2 years)t+1 a 0:73

(0:27)�0:17(�1:12)

�0:40(�2:78)

0:099

b �0:01(�0:71)

0:01(0:24)

0:32(2:54)

�0:34(�2:41)

0:075

c �0:01(�0:65)

0:02(0:47)

0:34(2:64)

�0:54(�4:03)

0:133

d �0:01(�0:90)

0:01(0:33)

0:36(2:68)

�0:22(�1:42)

�0:45(�3:35)

0:154

dy(3 years)t+1 a �0:30

(�0:14)�0:09(�0:67)

�0:44(�3:18)

0:087

b �0:01(0:72)

0:01(0:26)

0:34(2:61)

�0:27(�2:08)

0:053

c �0:01(�0:78)

0:02(0:45)

0:37(2:69)

�0:55(�3:93)

0:154

d �0:01(�0:92)

0:02(0:36)

0:38(2:73)

�0:13(�0:99)

�0:49(�3:59)

0:154

dy(4 years)t+1 a �0:60

(�0:33)�0:03(�0:28)

�0:46(�3:37)

0:089

b �0:01(�0:64)

0:02(0:42)

0:32(2:70)

�0:21(�1:80)

0:038

c �0:01(�0:76)

0:02(0:62)

0:36(2:80)

�0:53(�3:73)

0:165

d �0:01(�0:83)

0:02(0:56)

0:37(2:84)

�0:07(�0:61)

�0:50(�3:65)

0:157

dy(5 years)t+1 a �0:78

(�0:46)0:00(0:02)

�0:46(�3:49)

0:096

b �0:01(�0:55)

0:03(0:62)

0:29(2:76)

�0:17(�1:58)

0:026

c �0:01(�0:71)

0:03(0:84)

0:33(2:93)

�0:51(�3:54)

0:169

d �0:01(�0:73)

0:03(0:79)

0:33(2:95)

�0:03(�0:30)

�0:50(�3:66)

0:158

dy(10 years)t+1 a �1:35

(�0:89)0:05(0:70)

�0:44(�3:71)

0:121

b �0:00(�0:33)

0:06(1:57)

0:09(1:09)

�0:07(�0:78)

0:000

c �0:01(�0:54)

0:06(1:99)

0:13(1:90)

�0:43(�2:91)

0:138

d �0:01(�0:49)

0:06(1:95)

0:13(1:78)

0:05(0:60)

�0:45(�3:42)

0:129

50

Table 9Out-of-Sample predictions of monthly yield changes - di¤erent dealer groups

Test statistics in bold are signi�cant at the 5 percent level or betterForecast sample Comparison MSEU=MSER Test statistic 95% cv

dy(1Y )t+1

Aug. 01- Sep.05 OF1 vs const 0.90 5.49 1.47

" OF2 vs const 0.91 5.01 1.47

" OF2 vs const +OF1 0.98 1.24 1.47

dy(2Y )t+1

Aug. 01- Sep.05 OF1 vs const 0.95 2.72 1.47

" OF2 vs const 0.89 6.49 1.47

" OF2 vs const +OF1 0.94 3.27 1.47

dy(3Y )t+1

Aug. 01- Sep.05 OF1 vs const 0.98 1.04 1.47

" OF2 vs const 0.89 6.68 1.47

" OF2 vs const + OF1 0.93 4.15 1.47

dy(4Y )t+1

Aug. 01- Sep.05 OF1 vs const 0.99 0.28 1.47

" OF2 vs const 0.89 6.66 1.47

" OF2 vs const + OF1 0.92 4.68 1.47

dy(5Y )t+1

Aug. 01- Sep.05 OF1 vs const 1.00 -0.10 1.47

" OF2 vs const 0.89 6.67 1.47

" OF2 vs const + OF1 0.91 5.07 1.47

dy(10Y )t+1

Aug. 01- Sep.05 OF1 vs const 1.02 -0.98 1.47

" OF2 vs const 0.89 6.30 1.47

" OF2 vs const + OF1 0.90 5.88 1.47

51

Table 10In-sample predictions of monthly yield changes based on large vs. small dealers

All coe¢ cients are mulitplied by 100 and are in bold when signi�cant at the 5 percent level or

better

Maturity model FSt F1t F2t F3t OF3t OF4t Adj:R2

dy(1 year)t+1 a 8:77

(2:07)�0:15(�1:27)

�0:33(�2:53)

0:183

b �0:01(�0:68)

0:05(1:12)

0:24(2:53)

�0:38(�3:67)

0:095

c �0:00(�0:21)

0:03(0:73)

0:22(2:40)

�0:44(�3:38)

0:110

d �0:01(�0:59)

0:01(0:13)

0:27(2:58)

�0:29(�2:15)

�0:35(�2:86)

0:163

dy(2 years)t+1 a 1:06

(0:39)�0:12(�0:98)

�0:44(�4:04)

0:115

b �0:01(�0:91)

0:02(0:47)

0:32(2:55)

�0:31(�3:10)

0:055

c �0:00(�0:18)

0:00(0:05)

0:32(2:64)

�0:52(�4:55)

0:148

d �0:01(�0:59)

0:01(0:13)

0:35(2:69)

�0:19(�1:61)

�0:46(�4:09)

0:160

dy(3 years)t+1 a �0:10

(�0:05)�0:04(�0:41)

�0:47(�4:26)

0:107

b �0:01(0:84)

0:02(0:45)

0:34(2:57)

�0:24(�2:29)

0:038

c �0:00(�0:30)

0:00(0:02)

0:35(2:74)

�0:52(�4:50)

0:166

d �0:01(�0:50)

0:00(0:06)

0:37(2:75)

�0:11(�0:97)

�0:49(�4:16)

0:162

dy(4 years)t+1 a �0:47

(�0:25)0:00(0:00)

�0:47(�4:00)

0:108

b �0:01(�0:73)

0:02(0:58)

0:32(2:63)

�0:18(�1:74)

0:028

c �0:00(�0:29)

0:01(0:18)

0:34(2:89)

�0:50(�4:08)

0:171

d �0:00(�0:38)

0:01(0:21)

0:35(2:86)

�0:06(�0:51)

�0:48(�3:86)

0:161

dy(5 years)t+1 a �0:68

(�0:39)0:02(0:23)

�0:45(�3:74)

0:111

b �0:01(�0:62)

0:03(0:75)

0:29(2:67)

�0:15(�1:40)

0:019

c �0:00(�0:24)

0:01(0:39)

0:31(3:03)

�0:50(�3:72)

0:168

d �0:00(�0:27)

0:01(0:41)

0:32(2:98)

�0:03(�0:24)

�0:46(�3:58)

0:157

dy(10 years)t+1 a �1:28

(�0:81)0:03(0:29)

�0:38(�2:87)

0:106

b �0:00(�0:39)

0:06(1:66)

0:09(1:09)

�0:08(�0:74)

0:000

c �0:00(�0:12)

0:06(1:52)

0:11(1:74)

�0:36(�2:65)

0:109

d �0:00(�0:08)

0:05(1:54)

0:11(1:63)

0:01(0:11)

�0:37(�2:60)

0:096

52

Table 11Out-of-Sample predictions of monthly yield changes - di¤erent dealer groups

Test statistics in bold are signi�cant at the 5 percent level or betterForecast sample Comparison MSEU=MSER Test statistic 95% cv

dy(1Y )t+1

Aug. 01- Sep.05 OF3 vs const 0.93 3.85 1.47

" OF4 vs const 0.90 5.83 1.47

" OF4 vs const +OF3 0.93 4.00 1.47

dy(2Y )t+1

Aug. 01- Sep.05 OF3 vs const 0.98 1.09 1.47

" OF4 vs const 0.85 8.84 1.47

" OF4 vs const +OF3 0.88 7.31 1.47

dy(3Y )t+1

Aug. 01- Sep.05 OF3 vs const 1.00 -0.14 1.47

" OF4 vs const 0.85 9.17 1.47

" OF4 vs const + OF3 0.86 8.35 1.47

dy(4Y )t+1

Aug. 01- Sep.05 OF3 vs const 1.01 -0.61 1.47

" OF4 vs const 0.85 8.87 1.47

" OF4 vs const + OF3 0.86 8.61 1.47

dy(5Y )t+1

Aug. 01- Sep.05 OF3 vs const 1.02 -0.83 1.47

" OF4 vs const 0.86 8.56 1.47

" OF4 vs const + OF3 0.86 8.64 1.47

dy(10Y )t+1

Aug. 01- Sep.05 OF3 vs const 1.02 -1.24 1.47

" OF4 vs const 0.89 6.29 1.47

" OF4 vs const + OF3 0.89 6.63 1.47

53

Table 12In-sample predictions of monthly yield changes based on share of delayed

publication tradesAll coe¢ cients are mulitplied by 100 and in bold when signi�cant at the 5 percent level or

better. A coe¢ cient value of 1 indicates a change of 100 basis points in the yield

Maturity model FSt F1t F2t F3t OF5t OF6t Adj:R2

dy(1 year)t+1 a 8:22

(1:87)�0:06(�0:52)

�0:50(�2:48)

0:206

b �0:01(�0:73)

0:05(1:17)

0:27(2:62)

�0:38(�2:83)

0:101

c 0:01(0:38)

0:02(0:49)

0:20(2:19)

�0:65(�4:82)

0:167

d �0:00(�0:07)

0:03(0:69)

0:24(2:34)

�0:19(�1:12)

�0:52(�2:83)

0:178

dy(2 years)t+1 a 0:69

(0:27)�0:04(�0:28)

�0:65(�4:50)

0:143

b �0:01(�1:01)

0:03(0:53)

0:35(2:65)

�0:35(�2:61)

0:078

c �0:00(�0:04)

�0:01(�0:23)

0:30(2:40)

�0:71(�5:29)

0:179

d �0:00(�0:33)

�0:00(�0:10)

0:32(2:42)

�0:12(�0:79)

�0:62(�4:32)

0:176

dy(3 years)t+1 a �0:36

(�0:18)0:01(0:05)

�0:65(�4:94)

0:125

b �0:01(�0:98)

0:02(0:51)

0:37(2:68)

�0:30(�2:38)

0:066

c �0:00(�0:18)

�0:01(�0:24)

0:32(2:52)

�0:67(�4:71)

0:174

d �0:00(�0:32)

�0:01(�0:16)

0:34(2:47)

�0:07(�0:51)

�0:62(�4:33)

0:165

dy(4 years)t+1 a �0:68

(�0:36)0:03(0:24)

�0:61(�4:35)

0:114

b �0:01(�0:88)

0:03(0:64)

0:35(2:77)

�0:26(�2:19)

0:058

c �0:00(�0:18)

�0:00(�0:05)

0:31(2:66)

�0:61 0:164

d �0:00(�0:26)

0:00(0:01)

0:32(2:57)

�0:05(�0:34)

�0:58(�3:70)

0:153

dy(5 years)t+1 a �0:86

(�0:49)0:04(0:31)

�0:57(�3:84)

0:109

b �0:01(�0:78)

0:04(0:82)

0:32(2:86)

�0:23(�2:03)

0:050

c �0:00(�0:14)

0:01(0:19)

0:29(2:79)

�0:56(�3:60)

0:152

d �0:00(�0:20)

0:01(0:24)

0:29(2:66)

�0:03(�0:25)

�0:53(�3:23)

0:140

dy(10 years)t+1 a �1:42

(�0:89)0:01(0:05)

�0:42(�2:40)

0:081

b �0:01(�0:56)

0:06(1:78)

0:12(1:56)

�0:17(�1:49)

0:011

c �0:00(�0:06)

0:04(1:30)

0:09(1:37)

�0:39(�2:31)

0:076

d �0:00(�0:12)

0:04(1:39)

0:10(1:32)

�0:03(0:23)

�0:37(�1:91)

0:063

54

Table 13Out-of-Sample predictions of monthly yield changes - di¤erent dealer groups

-delayed publication of tradesTest statistics in bold are signi�cant at the 5 percent level or better

Forecast sample Comparison MSEU=MSER Test statistic 95% cv

dy(1Y )t+1

Aug. 01- Sep.05 OF5 vs const 0.93 3.72 1.47

" OF6 vs const 0.83 10.89 1.47

" OF6 vs const +OF5 0.89 6.32 1.47

dy(2Y )t+1

Aug. 01- Sep.05 OF5 vs const 0.96 1.97 1.47

" OF6 vs const 0.82 11.50 1.47

" OF6 vs const +OF5 0.87 8.02 1.47

dy(3Y )t+1

Aug. 01- Sep.05 OF5 vs const 0.98 1.00 1.47

" OF6 vs const 0.84 10.27 1.47

" OF6 vs const + OF5 0.87 7.91 1.47

dy(4Y )t+1

Aug. 01- Sep.05 OF5 vs const 0.99 0.55 1.47

" OF6 vs const 0.85 9.20 1.47

" OF6 vs const + OF5 0.87 7.46 1.47

dy(5Y )t+1

Aug. 01- Sep.05 OF5 vs const 0.99 0.33 1.47

" OF6 vs const 0.86 8.46 1.47

" OF6 vs const + OF5 0.88 6.98 1.47

dy(10Y )t+1

Aug. 01- Sep.05 OF5 vs const 1.00 0.01 1.47

" OF6 vs const 0.92 4.63 1.47

" OF6 vs const + OF5 0.94 3.08 1.47

55

2 Graphs

56

Figure 1: The predictive power of order �ow on daily 3 year changes vs RW

400 600 800 1000 1200 1400­0.000001

0.000000

0.000001

0.000002

0.000003

0.000004

0.000005

Figure 2: The predictive power of the three pricipal components of forward rates on daily 3year changes vs RW

400 600 800 1000 1200 1400­0.000005

­0.000004

­0.000003

­0.000002

­0.000001

­0.000000

0.000001

2.1 Tables and graphs

57

Figure 3: The predictive power of order �ow and the three pricipal components of forward rateson daily 3 year changes vs RW

400 600 800 1000 1200 1400­0.000001

0.000000

0.000001

0.000002

0.000003

0.000004

0.000005

Figure 4: Predicting 3 year monthly yield changes out-of-sample with short and medium termorder �ow vs RW

24 26 28 30 32 34 36 38 40 42 44 46 48 50 52 54 56 58 60 62 64 66 68 70 72 74 76­0.0001

0.0000

0.0001

0.0002

0.0003

0.0004

0.0005

0.0006

58

Figure 5: Predicting 3 year monthly yield changes out-of-sample with the third common factorin forward rates vs RW

24 26 28 30 32 34 36 38 40 42 44 46 48 50 52 54 56 58 60 62 64 66 68 70 72 74 76­0.00005

0.00000

0.00005

0.00010

0.00015

0.00020

0.00025

0.00030

0.00035

Figure 6: Predicting 3 year monthly yield changes out-of-sample with short and medium termorder �ow and the third forward rate factor vs RW

24 26 28 30 32 34 36 38 40 42 44 46 48 50 52 54 56 58 60 62 64 66 68 70 72 74 76­0.0002

0.0000

0.0002

0.0004

0.0006

0.0008

0.0010

0.0012

Figure 7:

59

Figure 8: Predicting 2 year yield changes - predictive power of order �ow of dealers of smallshare of delayed publication trades

25 30 35 40 45 50 55 60 65 70 75­0.00001

0.00000

0.00001

0.00002

0.00003

0.00004

0.00005

0.00006

Figure 9: Predicting 2 year yield changes - predictive power of order �ow of dealers with largeshare of delayed publlication trades

25 30 35 40 45 50 55 60 65 70 750.00000

0.00002

0.00004

0.00006

0.00008

0.00010

Figure 10:

60

Figure 11: Predicting 2 year yield changes - additional predictive power when adding order�ow of dealers with large share of delayed publication trades to the order �ow of dealers withsmall share

25 30 35 40 45 50 55 60 65 70 750.00000

0.00001

0.00002

0.00003

0.00004

0.00005

0.00006

0.00007

61