Agent-based social simulation: a method for assessing the ...web.csag.uct.ac.za/~gina/Gina_Ziervogels...Agent-based social simulation: a method for assessing the impact of seasonal

AGRICULTURALSYSTEMS

www.elsevier.com/locate/agsy

Agricultural Systems 83 (2005) 1–26

Agent-based social simulation: a methodfor assessing the impact of seasonalclimate forecast applications among

smallholder farmers

Gina Ziervogel a,*, Mike Bithell b, Richard Washington c,Tom Downing a

a Stockholm Environment Institute Oxford Office, 10B Littlegate Street, Oxford OX1 1QT, UKb Department of Geography, University of Cambridge, Downing Place, Cambridge CB2 3EN, UK

c School of Geography and the Environment, 1a Mansfield Road, University of Oxford, OX1 3TB, UK

Received 3 January 2003; received in revised form 10 February 2004; accepted 11 February 2004

Abstract

Seasonal climate forecasts provide probabilistic information on future climate on time-

scales of two to three months. Where this information is not presently used it is difficult to

evaluate the impact it might have. In order to justify disseminating the information to mar-

ginal groups it is important that the potential impact of the forecast is explored so that the

negative and positive effects are at least partially appreciated before use of the information

is widely promoted. We use an agent-based social simulation model, based on empirical evi-

dence from field work in Lesotho, to assess the impact of using seasonal forecasts among

smallholder farmers. The impact of using the forecast depends on the agents’ initial household

characteristics, what options they choose in responding to the forecast and the trust they place

in the forecast (which in turn depends on their ability to learn and to follow their neighbours).

Interaction of climate, crop productivity and social factors determines how much household-

agents benefit or lose, evaluated in terms of crop yields and likelihood of exhausting food

storage. Adoption of the forecast has the potential to decrease starvation among marginal

household-agents but poor forecasts may do more harm than good. This work suggests that

if forecasts are not correct more than 60–70% of the time, then they are unlikely to benefit

* Corresponding author. Present address: Climate Systems Analysis Group, Department of Environ-

mental and Geographical Science, University of Cape Town, Private Bag, Rondebosch 7701, South

Africa. Tel.: +27-21-650-2999; fax: +27-21-650-773.

E-mail addresses: [email protected] (G. Ziervogel).

0308-521X/$ - see front matter � 2004 Elsevier Ltd. All rights reserved.

doi:10.1016/j.agsy.2004.02.009

mail to: [email protected]

2 G. Ziervogel et al. = Agricultural Systems 83 (2005) 1–26

poor farmers. Poor forecasts, or forecasts that fail badly, when they do fail, lead to longer

adoption timescales for forecast use. Further investigation into the impact of the forecast at

the village level is encouraged before dissemination is actively pursued without appreciating

potential impacts.

� 2004 Elsevier Ltd. All rights reserved.

Keywords: Seasonal climate forecast applications; Agent-based social simulation; Marginal farmers;

Lesotho; Climate adaptation

1. Introduction

International efforts to improve societal responses to seasonal climate forecasts

are a priority because of their potential for ameliorating the impacts of climate var-

iability on marginal groups whose livelihoods are often highly impacted by climate

(Agrawala et al., 2001; Hansen, 2002; Anderson, 2003). Despite this priority, thereare few examples of how forecasts have been used by marginal groups (Eakin,

2000; Vogel, 2000). This dearth of information leads to uncertainty as to how fore-

cast development should be approached. The understanding of the impacts that fore-

casts might have on marginal groups is paramount to evaluating adaptation to

climate variability, a growing concern for the climate change community (Smit

et al., 2000).

Research has shown that marginal farmers (those individuals that struggle to se-

cure access to resources and services) are interested in receiving the seasonal fore-cast but the communication systems responsible for disseminating the forecast are

poorly developed and resource and cultural constraints restrict response options

(Archer, 2003; Eakin, 2000; Kirshen and Flitcroft, 2000; Mukhala et al., 2000;

O’Brien et al., 2000; Vogel, 2000; Phillips et al., 2001; Walker et al., 2001; Ziervo-

gel, 2004; Ingram et al., 2002; Patt and Gwata, 2002). Because forecasts are not

widely used at present, the impacts of forecasts are difficult to assess (Ziervogel

and Calder, 2003). Modelling approaches have been used successfully to assess

the value of using seasonal forecasts for tactical decisions in other counties, suchas Australia (Carberry et al., 2000; Hammer et al., 1996), but it is necessary to in-

clude the possible decision options when dealing with a group that has not previ-

ously used forecasts (Gadgil et al., 2002). Hansen (2002) suggests that a

combination of modelling and descriptive approaches can help to provide holistic

analyses and a co-learning approach where farmers and research both benefit.

Other research has suggested that a modelling approach that incorporates climate

and crop constraints and socio-economic factors be used as a tool for assessing

the potential value and use of forecasts in southern Africa (Vogel, 2000). The modelpresented in this paper combines qualitative decision characteristics with quantita-

tive environmental information and so integrates descriptive and modelling

approaches.

The model represents household response to the integration of forecasts into rural

communities at the village level. The objective is to develop a simulation that illus-

G. Ziervogel et al. = Agricultural Systems 83 (2005) 1–26 3

trates prospective pathways that might occur if households integrate the forecast into

their decisions. Specifically we aim to investigate:

1. The pay-offs of using the forecast as a function of forecast accuracy.

2. How mean trust in the forecast of farmer-agents grows or declines as a function of

forecast accuracy.3. What the benefit of using the forecast might be in terms of yield, with and without

interactions among farmer-agents and for different wealth profiles.

Forecast accuracy here refers specifically to rainfall, and is the percentage of time

that the forecast and actual rainfall are in agreement. The modelled behaviour is sup-

ported by empirical data from Lesotho, giving a baseline data set with a realistic dis-

tribution of household characteristics and response options for using forecasts. The

model is an abstraction of the Lesotho case study, but it does allow the analyst to

explore the potential of seasonal forecasts among smallholder farmers. Since farmershave not received and used climate forecasts, the model is one way to explore scenar-

ios of future development.

Section 2 describes the seasonal forecasts and highlights the difficulties associated

with assessing their usefulness. The case study of smallholder farmers in Lesotho, on

which the model is based, is then discussed. Agent-Based Social Simulation (ABSS)

is then introduced as a technique that allows crop, climate and social factors to be

coupled together for the exploration of the impact of seasonal forecast use. Section

3 outlines the model development and describes the results obtained; Section 4discusses possible extensions of the model and the final section concludes with a

model evaluation and the implication of the findings for future forecast application

developments.

2. Background: seasonal forecasts, field work and agent-based social simulation models

2.1. Forecast dissemination and adoption

During the past few decades, climatologists have improved their ability to predict

the seasonal climate (Cane et al., 1986; Palmer and Anderson, 1994; Martin et al.,

2000; Murphy et al., 2001). These seasonal climate forecasts are based on ocean–

atmosphere interaction such that sea surface temperatures (SSTs) determine future

atmospheric perturbation states (Washington and Downing, 1999). This predictabil-

ity is largely confined to the tropical atmospheric circulation and is principally de-

pendent on El Ni~no-Southern Oscillation (ENSO)-related anomalies (Masonet al., 1996). Seasonal climate forecasts are available for temperature and rainfall.

Regional forecasts are produced in a number of regions including Australia, Latin

America, the Caribbean, Pacific islands, Southeast Asia, West Africa, the Greater

Horn of Africa and southern Africa (Basher et al., 2001; Stone and de Hoedt,

2000). In southern Africa, where the rainy season stretches from October to March,

a regional consensus meeting, organised by the Drought Monitoring Centre (DMC)-

Harare, is held in September (O’Brien et al., 2000; Basher et al., 2001). A six-month

consensus forecast, that integrates the output of numerous seasonal forecast models,


is issued for October–March. A mid-season correction meeting is held in December

to update the forecast for January–February–March. Each country adapts the re-

gional forecast using local data to produce a national forecast.

The nature of the seasonal forecast means that it is not appropriate for all users,

as there are numerous sector-specific constraints associated with using the forecast,which can limit uptake (Nicholls, 1999; Orlove and Tosteson, 1999; Patt, 2001).

These can be separated broadly into problems of forecast dissemination and prob-

lems of forecast use. The national meteorological organizations are responsible for

disseminating the forecast within their country. Most countries in southern Africa

do not have well-developed dissemination strategies and so there is not great aware-

ness of the forecast at the national level and even less at the local level (O’Brien et al.,

2000). Furthermore, the coarse scale and temporal resolution of the forecast make it

difficult to apply locally. Even with effective distribution of the forecast, the level ofuse is dependent on further factors; both the skill of the forecast and the credibility

of the source play a role. The skill of the forecast is a measure of the degree of cor-

respondence between forecasts and observation (Murphy, 1997). Skill scores such as

this are an attempt to summarise the joint probability distribution of many climate

variables, and as such may not bear directly on the probability of rainfall occurring,

for example. The credibility of the forecast, which will determine the amount of trust

users place in the forecast, is partly determined by past skill and partly by forecast

communicators’ reputations (Patt, 2000, 2001; Patt and Gwata, 2002). Trust will alsobe affected by the way the forecast supports or contradicts local beliefs about the cli-

mate (Eakin, 1999). These beliefs might be based on cultural norms or environmental

indicators. Improved credibility may therefore depend on improved forecasts and

better data, it may depend on better communication that emphasises the probabilis-

tic nature of the forecast and comes from a respected source or it may depend on

adjustment of cultural and social perceptions that view the forecast in a favourable

light.

Forecasts are innovations that have not been widely adopted, particularly amongmarginal groups. Traditional information dissemination literature would assess the

success of adoption of forecasts, a type of innovation, as a function of the numbers

of users that have already adopted the innovation in comparison to those who have

not (Rogers, 1995). This may be inappropriate when there is not a high adoption rate

and so few data. It does not enable the stochastic nature of adoption, associated with

prospective technological innovations, to be anticipated (Mazzocco et al., 1992). As-

sessing the possible use and impact of forecasts, when they have not been available

previously, requires a process-based model for the exploration of available futurepathways of development. In this way, the potential range of impacts of forecast

use can be captured (Stern and Easterling, 1999; Downing et al., 2000).

2.2. Lesotho: case study

Lesotho, a small mountainous country in southern Africa, provides a natural lab-

oratory in which to assess how seasonal forecasts are being disseminated, perceived

and used at a national and local scale. In Lesotho, the seasonal forecast is developed


using statistical methods that combine past climate data with seasonal ocean and at-

mospheric data. An analysis of the 2000 October–November–December rainfall

forecast showed a skill of 56% of a hindcast compared to past data (Peshoane,

2000). The systems that disseminate forecasts in Lesotho are not particularly well de-

veloped (Ziervogel and Downing, in press). Lesotho Meteorological Services (LMS)has issued the forecast for six years but dissemination is not widespread. LMS hold

an annual workshop to announce the forecast but the meeting is attended primarily

by government officials, who tend not to disseminate the forecast beyond their im-

mediate colleagues. The radio programmes that report on the workshop do not im-

prove awareness of the forecast much, as they read out a press statement rather than

expanding on the forecast characteristics and possible uses.

Field work, undertaken in one village in southern Lesotho, Ha Tlhaku, and sup-

ported by field work from other parts of Lesotho, has been used to develop themodel. Modelled requirements for food, available resources and likely reactions to

introduction of the forecast have been based on field work undertaken between

September 2000 and November 2001.

A range of data elicitation techniques such as surveys, workshops and participa-

tory tools were used to gather data. A role-play exercise was developed to elicit how

farmers might use forecast information that they had not heard before. This partic-

ipatory method enabled farmers to think through the situation with which they were

faced at the beginning of the growing season, when presented with new informationabout the climate (the seasonal forecast). They were asked whether they would in-

tegrate the forecast into their decisions and if so, how. The scenario was first pre-

sented for the present year when they were asked about what decisions they faced

that season. It was then repeated for the following year when a forecast of below

normal rainfall was expected and lastly for a year with a forecast for above normal

rainfall.

The role-play exercise elicited options that farmers thought they might pursue if

they received the forecast, as seen in Table 1. The options suggest that abundant re-sources are not required for all responses. Some suggestions were more prevalent

than others and some more appropriate for a wide range of users. Options such as

changing cropping densities can be pursued by all households and many can change

crop type by bartering with their neighbours, as is a common practise. Changing to

drought resistant varieties might only be an option for households who can afford to

buy new seed, unlike most households that keep their seed from the harvest to plant

the following year.

The types of decisions that people mentioned they might undertake in response tothe forecast were mostly short-term tactical decisions such as sowing less maize and

wheat, planting earlier or protecting the livestock, rather than long-term strategic de-

cisions. These decisions are a form of agricultural risk management. The focus on

short-term decisions is expected when using new information. When these decisions

seem an appropriate way of responding to the forecast, then more long-term deci-

sions can be made in response to the forecast, depending on the risk preference of

the decision maker (Carberry et al., 2000). The decisions chosen for use in the model,

in response to the forecast, are the adjustment of cropping densities and the ratio of

Table 1

Suggestions on how to adapt to below and above normal rainfall as established in the role-play exercise

Advice given to other farmers about what might be done if:

Predictions for below normal rainfall Predictions for above normal rainfall

Plant drought resistant crops Grow more vegetables as cash crops

Sow less maize and wheat Plant earlier

Reduce density of field crops Increase sharecropping (where one party supplies

Irrigate if the streams still have water: dig furrows

next to fields to stop wilting

the land and the other the inputs)

Increase crop densities

Collect water and pour over plants Plant different crops to my neighbour so that

Plough lucerne and teff as animal feed since they

survive harsh conditions

there is a market

Be aware of diseases and pests for crops and

Reduce the number of animals which cannot

survive the harsh conditions (by selling)

animals

Practise winter breeding since there will be water

Take animals to the mofebong (rangelands) where

there is better grazing and water

and feed for the sheep and goats

Build animal shelters


maize to sorghum planted. These decisions are generally available to a wide range of

farmers, including those with few resources.

The role-play results also suggest that farmers are in practice liable to ignore fore-

casts of dry conditions, but take heed of those for good rains. They do not want to

accept that the season might be unfavourable, as that could result in food deficits

and they do not have many alternative off-farm strategies. Additionally, farmers

are more likely to use the forecast when they had been exposed to it for a numberof years (Ziervogel, 2004).

Table 2 presents a selection of household characteristics, gathered from household

interviews in Ha Tlhaku, that shows the diversity of field numbers and size (repre-

sented by amount of seed needed for one field as individuals did not always know

Table 2

Household characteristics with regards to field, yield and food needs based on field work data

Household

(HH)

identification

Number

of fields

Amount

of seed for

all fields

(kg/year)

Yield in best

conditions

(kg/year)

Yield in worst

conditions

(kg/year)

Number

of HH

members

Maize

consumed

by HH in

1 month (kg)

1 1 12.5 240 80 3 25

2 1 12 600 25 7 70

3 2 45 500 50 6 80

4 3 60 960 160 8 75

5 3 100 1920 480 7 80

Note. This is a selection of data that show a spread of household characteristics, as based on interviews

in Ha Tlhaku. Although many households were interviewed in different villages, these data are extracted

from the 11 household profiles that were compiled in Ha Tlhaku. More detailed data, on Ha Tlhaku and

other villages, can be obtained from the authors.


the areal extent) and the yield in good and bad conditions. Most households in Le-

sotho have a vegetable garden and one field, with the average field size being 1.4 ha

(Gay and Hall, 2000). Although the range in data is quite wide, household typologies

have been distilled from it based on livelihood characteristics and have been used to

characterise the farmer-agents in the ABSS model.

2.3. Agent-based social simulation

ABSS is a computational technique that attempts to model human behaviour. Peo-

ple, households or larger institutions are represented by a set of logical rules that deter-

mine uniquely for each agent how they will act in a given environment (Edmonds et al.,

1996; Chattoe, 2000). Although these rules, and the priorities of agents, are stated, the

inclusion of non-linear variables, and the processes of interaction between agents, andbetween agents and their environment, allows novel behaviour to emerge.

The agents can be said to hold opinions (beliefs about the environment or other

agents based on limited knowledge), and these opinions may come to be held by an

entire community through communication. Psychological, cultural and political con-

straints can be included as determinants of how agents behave. The contextual pa-

rameters can be specified and changed, which enables controlled experiments on

the effect of changing social, economic and physical parameters to be measured

(Moss, 1999). Non-economic motivations, such as preservation of environmentalor cultural values, can enter into decisions, an important advantage over purely eco-

nomically motivated models. This enables economic, social and biophysical elements

to be considered in one model (Miller, 1998; Blench, 1999; Washington and Down-

ing, 1999; Phillips et al., 2001). ABSS, therefore enables adaptation options to be

evaluated; a growing concern for the climate change community (Smit et al.,

2000). At the same time, the agents can have an internal metabolism and require-

ments for food, shelter and other necessities. Rules both for farming activity and

for economic behaviour govern the ability of individual agents to acquire these ne-cessities.

ABSS facilitates anticipatory evaluation of technology adoption. It is an explor-

atory method consistent with social science interests, as well as representing social

behaviour in formal ways that link to models of natural systems. In exploratory

models, dynamic behaviour is emphasised rather than equilibrium, which enables

complex systems to be opened up so that rather than predicting outcomes, counter-

intuitive behaviour can explored (Barreteau et al., 2001). ABSS has not been applied

previously to seasonal forecasting, although other models have been used toaddress the anticipatory aspects of forecasts as emerging technology (Mazzocco

et al., 1992).

3. The model

The set of simulated farmers used in the model is based on the information

described in Section 2.2. The farmer characteristics include size of household, food


requirements, rate of change of trust in the forecast and the household characteristics

from Table 2. A summary of the agents in the model and relationships between them

is presented in Fig. 1.

Farmer-agents respond differently to the forecast depending on their initial re-

source base and decision-making preferences. Table 3 shows the progression ofthe model from the representation of crop, climate and forecast, through responses

of individual farmers, leading finally to a set of 700 interacting household-agents,

representative of a medium sized village.

3.1. Climate and crop parameterization

The climate data used in the model are based on the January–February–March

(JFM) growing season rainfall data for Lesotho (Peshoane, 2000). The distributionof the JFM rainfall data for 39 years from the period 1960 to 2000 (no data were

available for 1986/1987), presented in Fig. 3, shows that while dry and ‘‘normal’’

years occur with near equal frequency, wet years are rather less probable. As the

forecast can only take one of the three values (above normal, normal or below nor-

Met. Office

Forecast Skill

The skill defines theprobability that theforecast agrees withthe "real" weather

World

Four timesteps per year

"Real"Weather

Climate sequenceForecastwith errors

Available Resources

Climate

Rainfall Variability data

Rainfall data defines thestatistical distribution ofwet, normal and dry years

Weather

Seasonal RainfallAN,N or BNRainfall randomlyselected for this yearusing climate distribution

Fields

Maize or SorghumPlant DensityThe fields act as containersthat hold crop information

Crop Parametrization

Equations for yield

Crops grow for twoseasons and then canbe harvested and addedto household store

Crop Decisions

Household

Number of fieldsNumber of occupantsForecast TrustFood storesRules for AdaptationNumber of occupantsdetermine consumption.Planting decisionsdepend on history offorecast success

Fig. 1. A schematic representation of the agent-based social simulation model. Each agent type is repre-

sented by a box, in which the first field gives the agent type. The second field shows the data held by the

agent and the third gives an indication of the agent behaviour. Arrows show the flow of information be-

tween the agents. All the agents are embedded in a world object, which controls the time steps.

Table 3

Progression leading to final runs of the ABSS

Progression Description

Background data Data are needed to represent agents and their environments in the

simulation. Some of the data was gathered using qualitative methods,

such as participatory research and some using quantitative methods,

such as surveys and secondary climate (rainfall) and agricultural data

(crop yield in response to cropping density and rainfall variation).

Individual cost of using

information

A pay-off matrix provides the reward structure of crop suitability relative

to actual and forecasted rainfall, based on field and secondary data.

Yield is determined by comparing how cropping density and crop type

ratios fare in response to expected rainfall versus unexpected rainfall.

Mean trust A model of trust in the forecast for individual farmers is developed

dependent on how the cost of using information varies with forecast

accuracy.

Dispersal of benefit of using

the forecast

Farmer-agents have the same model of trust but there is a realistic

diversity of household classes, based on different initial stores of grain,

number of fields and number of household members. Use of the forecast,

therefore, affects households differently.

Benefit of using the forecast

depending on community

interactions

Use of the forecast is influenced by interaction among farmer-agents that

depends on social norms, specifically the propensity to pay attention to

the opinions of others.


mal), we represent the rainfall by assuming 40% of years are ‘‘below normal’’, 40%

‘‘normal’’ and 20% ‘‘above normal’’, with rainfall taking only one of the three dis-

crete values 170, 300 and 460 mm respectively.

The two crop types considered are maize and sorghum, the primary crops plantedin Lesotho. Yield is calculated as a function of rainfall and planting density. Accord-

ing to Rowland (1993), the response of sorghum is largely linear with rainfall, but

with slope increasing as the density of planting increases. At a constant rainfall level,

the yield varies approximately quadratically with density. The maximum of the qua-

dratic moves to increased density and yield with increasing rainfall. The behaviour of

maize is similar, but the slope of yield with rainfall is steeper, leading to better sor-

ghum yields in dry conditions, but superior maize yield at higher rainfall. Annual

household yields in Ha Tlhaku varied significantly, so the national average yieldsfor Lesotho have been used; near to 1 ton/ha for maize and 0.8 tonnes/ha for sor-

ghum (Chakela, 1999; Bureau of Statistics, and Food and Agriculture Organisation,

2001). An approximation that captures the above features gives the following yield in

tonnes/ha:-

for maize

ð0:05þ 0:051r100Þðd10 þ 3Þ � ð0:00875þ 0:000032r100Þðd10 þ 3Þ2; ð1Þ

ð0:05þ 0:027r100Þðd10 þ 3Þ � ð0:007þ 0:000017r100Þðd10 þ 3Þ2 þ 0:3; ð2Þ
for sorghum, where r100 is the seasonal rainfall in hundreds of millimetres and d10 isthe density of plants in tens of thousands per hectare (i.e., d10 ¼ 1 implies 10,000
plants/ha). These relationships were derived for seasonal rainfall between 200 and


700 mm, and crop densities from 10,000 to 120,000 plants/ha. In Fig. 2, we show

yield as a function of crop density for median above normal (AN), normal (N) and

below normal (BN) JFM rainfall for both maize and sorghum. Also shown is the

corresponding yield as a function of rainfall (see Whiteman, 1981 in Rowland, 1993).

Although the intraseasonal rainfall variability is high, it is not included here, as amore detailed crop model would be required to capture the effects on crop yield.

3.2. Run 1: pay-offs of using the forecast as a function of forecast accuracy

We now consider how farmer-agents respond to the available forecast without re-

gard to their neighbours’ behaviour. The best-case behaviour for ‘‘normal’’ condi-

0

0.5

1

1.5

2

2.5

2 4 6 8 10 12 14 16 18 20

Yie

ld (

tonn

es/h

ecta

re)

Plant Density (10000/hectare)(a)

0

0.5

1

1.5

2

2.5

1 2 3 4 5 6 7

Yie

ld (

tonn

es/h

ecta

re)

JFM rainfall (100mm)(b)

Fig. 2. (a) Crop yield as a function of plant density for normal (300 mm), below normal (170 mm) and

above normal (460 mm) January–February–March rainfall for maize (solid line, 300 mm; dashed line,

170 mm and dotted line, 460 mm) and sorghum (plus (+) 300 mm, cross (�) 170 mm and star (*) 460

mm) (based on Figs. 4.8 and 4.9 in Rowland (1993)). (b) Crop yield as a function of rainfall for two dif-

ferent plant densities. Maize: Solid line 10,000 plants/ha and dashed 100,000 plants/ha; Sorghum: plus (+)

10,000 plants/ha and cross (�) 100,000 plants/ha.

0

2

4

6

8

10

12

100 150 200 250 300 350 400 450 500 550 600

Fre

quen

cy (

no. o

f yea

rs)

JFM total rainfall (mm)

Fig. 3. Frequency distribution of 39 years of January–February–March rainfall for Lesotho.


tions with our above parameterization has farmer-agents planting 80,000 plants/ha.

If agents hedge their bets against dry conditions and satisfy their preference formaize by planting 60% maize and 40% sorghum, as suggested in the field work, then

the average return, allowing for the variation in rainfall, is 960 kg (Table 5, middle

row). If 100 kg of seed is saved for next year’s planting, then in a household of eight

people with a field of 1 ha, each person can eat just over 100 kg of grain per year (cf.

the mean of 128 kg of maize per person per year calculated from Table 2). This is not

enough to ensure that households will not run out of food. Farmers in Ha Tlhaku

suggested that grain can be stored for approximately 2 years, so the model farmers

discard any grain that is older than this. Given a run of bad years, grain stocks maybecome exhausted. In Fig. 4(a), we show one realization of 50 years of simulated

grain storage for a single eight-person household-agent where there is no variation

in planting behaviour. The rainfall changes randomly with the distribution described

in Section 3.1. Note that in 20 of the years the household is short of food.

In the role-play exercise in Ha Tlhaku, farmers provided a number of options for

responding to the forecast (see Table 1). The two response options used in the model

are a change in the ratio of maize to sorghum planted and the alteration of cropping

density (Table 4).The non-linearity of the crop response to planting density can lead to very severe

penalties for farmer-agents whose planting is not well matched to the climate. This

has implications for the effect of following a seasonal forecast that is not 100% ac-

curate. Table 5 shows the effects on crop yield of maize and sorghum when planting

at optimal densities for the given forecast. This pay-off matrix illustrates that the

farmer-agent can reduce risk in years where above normal rain is forecast by plant-

ing a mixture of maize and sorghum, but this has less effect than simply planting at

lower density. For example, in the case where above normal rain is forecast (460mm), an agent who plants the optimal density of maize will be expecting a crop of

Table 4

Forecast response options

Forecast Option 1 Option 2

Crop ratio (maize:sorghum) Cropping density (seeds/ha)

Above normal 80:20 120,000

Normal 60:40 80,000

Below normal 40:60 40,000

0

0.5

1

1.5

2

2.5

3

3.5

4

0 5 10 15 20 25 30 35 40 45 50

Sto

rage

(to

nnes

)

Year(a)

0

2

4

6

8

10

12

14

16

0 5 10 15 20 25 30 35 40 45 50

Cos

t (to

nnes

of g

rain

)

Year(b)

Fig. 4. (a) Simulated times series of grain storage. Crop production for a single climate sequence of 50

years with 20% chance above normal rainfall, 40% normal and 40% below normal. Farmer-agents are as-

sumed to plant 60% maize and 40% sorghum, at a density of 80,000 plants/ha. (b) Cumulative cost of poor

years. Extra costs incurred as a result of crop failure in three cases. Solid line: no forecast, dotted line:

forecasts always incorrect, plus (+): 65% correct forecast, and at most 10% two terciles out (i.e., wet fore-

cast when it is actually dry or vice versa). Farmer-agents are assumed to adjust planting to give maximum

yield when the forecast is correct.


Table 5

Rainfall-yield pay-off matrix for sorghum and maize

Forecast Actual rainfall

AN N BN

Maize Sorghum Maize Sorghum Maize Sorghum

AN 2.3 1.3 1.0 0.7 0.1 0.2

N 2.0 1.4 1.2 0.9 0.4 0.5

BN 1.6 1.2 1.0 0.9 0.5 0.6

Notes. Expected yield amounts in tonnes per hectare. Rainfall amounts based on JFM total rainfall of

AN ¼ 460 mm, N ¼ 300 mm and BN ¼ 170 mm. Data are based on Eqs. (1) and (2), that were derived

from Rowland (1993), and calibrated using the field work.


nearly 2 tonnes/ha. If the actual rainfall is below normal, the return will be 0.1 ton-

nes/ha (although the rather wide bins in rainfall make this low a return somewhat

unrealistic), whereas an agent that has planted expecting normal weather will still

get 0.4 tonnes/ha of maize. So, following forecasts of high rainfall is a much higher

risk strategy than following those for low rainfall, unless the forecast accuracy for

wet years is considerably better than for dry. On the other hand, following the fore-

cast in dry years confers only a small benefit over behaving as for normal weather.

In Fig. 4(b), we show three 50-year time sequences for a representative eight-mem-ber household-agent. In each of the three cases the climate sequence is the same, and

the household-agent plants its crops as if it believes the forecast to be correct, when it

is available. The middle line shows the case when there is no forecast – poor years

lead to a deficit of grain, and the figure shows the cumulative cost of these years

in terms of the number of tonnes of grain the household-agent would have to obtain

through other means. The upper line shows the results of the worst case, where the

forecast invariably fails, i.e., it predicts wet years (AN) when the actual rainfall is be-

low normal and dry (BN) when there is above normal rain. In years when the rainfallis normal (N) it predicts either AN or BN at random. The lower line gives the result

when the forecast is correct 65% of the time. Incorrect forecasts are assumed to be as

bad as possible (AN when the rainfall is really BN and vice versa, which is to say

incorrect by two terciles) at most 10% of the time.

Although Fig. 4(b) shows that poor forecasts can be damaging, it does not show

how often the forecast needs to be right in order for farmer-agents to benefit overall.

This is illustrated in Fig. 5. Five hundred random climate sequences each of 50 years

were generated for the eight-member household-agent. The figure shows the accumu-lated cost at the 50 year mark as a function of the percentage of correct forecasts.

Each point shows the mean and SD cost from the 500 sequences. The dotted line

is the no-forecast case: Points must lie below this line for the forecast to be of benefit

to the household-agent. The crosses show the case where, when a forecast fails, it is

only really poor (i.e., forecast for a wet season when it is actually dry and a dry sea-

son when it is actually wet) at most 10% of the time. The upper points show the case

when failed forecasts are always as bad as possible. From this we conclude: (1) The

forecast must be correct more than about 60–70% of the time to benefit these

-2

0

2

4

6

8

10

12

14

0 20 40 60 80 100

Cos

t (to

nnes

of g

rain

)

Percentage of correct forecasts

Fig. 5. Cost at year 50 as a function of forecast accuracy. Dotted line: no forecast (the error bar on the

open square shows the standard deviation about this line). Crosses (�): forecast wrong by two terciles at

most 10% of the time, plus (+): failed forecasts always as incorrect as possible (see text).


household-agents and (2) When the forecast fails, it must be very poor no more than

about 10% of the time.

3.3. Run 2: mean trust as a function of forecast accuracy

So far we have examined the consequences for a farmer-agent who blindly followsa forecast when it is issued, irrespective of the impact on their productivity. The role-

play exercise in Ha Tlhaku suggests that farmers become more likely to accept a fore-

cast as familiarity increases, and that forecasts of poor conditions tend to be ignored

(see Section 2.2). On the other hand, farmers are likely to lose confidence in the fore-

cast if it fails often, or if they suffer from poor returns as a result of following a fore-

cast of good conditions that turns out to be incorrect. As a model for this we allocate a

floating point number to each farmer-agent that represents their level of trust in the

forecast. The trust increases linearly by one unit per year whenever the forecast provescorrect and decreases by the same amount when it is not. After 3 years, the farmer-

agents have built up enough trust that they are prepared to begin using the forecast.

In Table 6 we show the assumed probability of a farmer-agent using the forecast at

given level of trust, depending on whether the forecast is AN or BN. These numbers

reflect increasing confidence in the forecast, but a tendency to ignore those for below

normal rainfall, as indicated in the role-play exercise. If trust is less than the threshold

of three units, the forecast is not used at all.

To simulate the effect of failed forecasts on crop production, we adopt a ‘once bit-ten, twice shy’ formulation. The farmer-agents that are using the forecast drop their

trust level back to the threshold value if a forecast for a wet year fails. If at the same

time the resulting poor crop leads to exhausted food stocks, the subsequent rate of

Table 6

Level of trust for using the forecast

Trust Forecast

AN (%) BN (%)

3 75 50

4 85 60

5 95 70

>6 100 80


increase in trust level is halved. If the forecast leads to exhaustion of food stocks asecond time, then the farmer-agent adopts a ‘never again’ stance and does not use the

forecast at all thereafter. Long-term trust in the forecast depends largely on the fre-

quency with which wet conditions are forecast, but dry conditions are experienced in

practice. In Fig. 6, we show the effects on the level of forecast trust for different levels

of forecast accuracy. The figure shows the mean level of trust as a function of time,

averaged over the same 500 climate sequences used previously, now run for 100

years, and the same household of eight, consuming 100 kg of grain/person/year.

If the forecast accuracy is sufficiently high then the mean trust can reach high lev-els, although it tends to saturate rather than increase without limit. There is a scatter

about the saturated level that remains high, but it increases less rapidly than the t1=2

behaviour that would be expected for a random walk (not shown). If the forecast

fails badly more than at most 10% of the time (i.e., the forecast is for a wet year when

it is dry and vice versa), then the mean trust level reaches a maximum, and thereafter

declines. At the 10% level, the trust saturates after about 30 years, i.e., the quality of

0

1

2

3

4

5

6

7

8

9

0 20 40 60 80 100

Tru

st le

vel

Year

Fig. 6. Growth of mean trust when poor forecasts damage the trust level. As the number of failed wet year

forecasts increases, the mean level of trust starts to decline. Note the long time scale over which this takes

place. Scatter about the mean is of the same order as the mean itself. Forecasts are assumed correct 60% of

the time. Solid line: at most 5% of failed forecasts incorrect by two terciles, dashed 10%, plus (+) 15%.


failed forecasts has a significant controlling effect on trust levels. Note that the time-

scale is much larger than the 3-year timescale on which the forecast begins to be used

in the role-play exercise and is dependent on the accuracy of the forecast rather than

the rate of increase of trust. If the accuracy with which the forecast is correct drops

below 60%, then the trust level just saturates (although it begins to decline afterabout 90 years). In each case, at the 100 year mark, the scatter in the trust across

climate sequences is of the same order as the mean, implying that in many cases

the household-agents give up using the forecast even when on average the mean trust

is positive.

3.4. Run 3: collective effects

We now ask how things might change when there is a range of behaviour andwealth in a population, and how this might be changed by social interactions within

such a setting. We restrict ourselves, in this section, to the case where the forecast is

correct 60% of the time and incorrect by two terciles (i.e., as bad as possible) at most

10% of the time.

We create a ‘‘village’’ of 700 household-agents, which are selected from one of the

three household classes, each with a different initial store of grain, number of fields

and household members. The household classes (better-off, average, poor) reflect the

typical distribution of wealth in Lesotho and are based on field work and literatureon Lesotho (Gay and Hall, 2000; Turner et al., 2001 – see Table 7). Separate to this,

each household has an intrinsic sensitivity to the forecast, and a tendency to pay at-

tention to the opinions of others.

For initial storage, we assume a Gaussian distribution, with variance of 0.2 ton-

nes, and mean as given in the table. We distribute the rate of increase of trust, again

using a Gaussian distribution, with mean 0.5 and variance 0.2. Thus, only those on

the more optimistic fringe of the distribution increase their trust at the rate of pre-

vious sections. We use the same distribution for the drop in trust that occurs whena forecast fails, although individual farmer-agents may have different values for their

rates of increase and decrease of trust. The ‘once bitten, twice shy’ behaviour is also

given a Gaussian spread – when the forecast for a wet year proves dry, farmer-agents

drop their trust to a mean level of 2.5 and variance again of 0.2.

Opinion formation is an active area of current research. No definitive models exist

that adequately describe how a social system will respond to environmental stimulus

Table 7

Household class descriptions

Class

Better-off Average Poor

Number of fields 2 1 1

Number of members 4 6 8

Mean initial store 3 2 1

% of population 20 40 40


or how interactions between individuals within the system affect that response. We

examine one possibility, where people adjust their own trust level by using a

weighted average of theirs and those of others (see, e.g., Hegselmann and Krause,

2002). Denoting the trust of individual i at time t by T , we use the following recur-

rence relation between timesteps t and t þ 1:

Tiðt þ 1Þ ¼ TiðtÞ þX

j

aijðT; tÞTjðtÞ þ BiðtÞ; i 6¼ j ð3Þ

where indices i and j run from 1 to N, the number of individuals in the village, and

bold face T denotes the entire vector of trust for the whole village. B denotes the

background rate of change of trust as described in previous sections (a function of

time, since it may be positive or negative in a given timestep, depending on the

forecast outcome and whether the farmer-agent is adapting planting accordingly).

The vector of weights, a, may be a function of time and of the trust vector itself. We

imagine that four times a year, each villager is able to sample the trust level of everyother village member. This might take place at village meetings, for example. In the

simplest case, suppose that the weights are constant, equal and independent of T. Ifthe weights are all positive and they sum to one (i.e., aij ¼ 1=ðN � 1Þ), then the trust

vector rapidly collapses to a narrow range near the mean of the no-interaction trust

distribution, despite the variation present in B. So, if individuals assign equal total

weight to their own trust level and that of the rest of the village, a consensus of trust

is reached over time.

To allow more freedom in choice of a, suppose that in any timestep an individ-ual villager adjusts the weight they give to the trust of another by a fixed additive

constant k (e.g., 0.1). We remove the other restrictions on a, except that it must

remain positive (negative values have unfortunate run-away consequences for this

particular model, although a more realistic model could include the notion that

some people may like to do the opposite of what others believe). We start from

a ¼ 0 and adjust the weights upward if another villager’s reaction to the forecast

proves in the previous time step to have been right and downward if they were

wrong. Now Eq. (3) is non-linear, but the effect is that the trust vector collapsesalmost to the linear case, with a timescale determined by k. Fig. 7 illustrates the

history of trust variance (for all household classes) for a single model run in the

cases with no social interaction and with the linear and non-linear interaction

cases. In the no interaction case, the trust variance across the population tends

to rise gradually (owing to the spread in sensitivity to forecasts) until forecast

failure (near year 70 for example). Social interaction severely limits the variance

in trust, despite the different rates at which trust in the forecast changes for each

individual.We can now discuss the effect of the forecast by reference to household class. A

control run was performed in which no use of the forecast was made, and then runs

with no social interaction, but including the forecast, as in previous sections, and fi-

nally a run with both forecast and non-linear social interaction. In the control case,

where no attention is paid to the forecast, the results are rather uniform: because we

have no variation within a class other than initial storage, the two-year storage limit

0

0.5

1

1.5

2

2.5

3

3.5

4

0 10 20 30 40 50 60 70 80 90 100

Tru

st V

aria

nce

Year

Fig. 7. Variance of trust with and without social interaction. Plus (+) no interaction, cross (�) linear

interaction and star (*) non-linear interaction.


tends to lead to a single history for each class after only a few years. Once we intro-

duce forecasts, then the spread in belief also spreads the cost history. Five hundredruns of the model were again made, each with 700 household-agents.

The following frequency plots show the effects of the forecast for the no-social-

interaction and the non-linear interaction cases. We show the case for eight-member

(poor) household-agents and ask whether the number of cases in which food short-

age occurs is changed by using the forecast. Fig. 8(a) shows the number of runs in

which use of the forecast led to an increase in the amount that must be spent to avoid

starvation. Only in a few of the runs was a large fraction of the population disadvan-

taged by forecast use. The maximum mean extra expenditure (over a single village)over 50 years came to 0.9� 0.5 tonne of grain (i.e., nearly a full year of consump-

tion). The effect of non-linear interactions was to reduce slightly the number of runs

in which there was a deficit relative to the control, but the number of runs in which

nearly the whole population was affected was increased – when the forecast behaves

poorly, the whole population is involved, because of the convergence of opinion in

the model.

Fig. 8(b) shows the frequency plot for those benefited by the forecasts. In these

cases, spending as a result of running out of food was reduced, relative to the con-trol, by using the forecast. The results are very similar to Fig. 8(a), reflecting the fact

that a 60% accurate forecast is only just worthwhile for these household-agents (cf.

Fig. 5).

Reductions in spending were modest, with a maximum mean gain of 1.4� 0.2 ton-

nes relative to the control. The effect of non-linear interaction was again to slightly

reduce numbers gaining overall, although in this case we did not see increases in the

number of runs where nearly all the population was affected. The results for six-

member household-agents (not shown) were similar to those for eight members,but the six-member household-agents meet starvation levels less often. Gains and

0

20

40

60

80

100

120

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Num

ber

of r

uns

Fraction of Population: 8 member Households(a)

0

10

20

30

40

50

60

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Num

ber

of r

uns

Fraction of Population: 8 member Households(b)

Fig. 8. (a) Number of runs in which a given fraction of the population spends more than control for eight-

member household-agents. Solid lines: no interaction and dashed: non-linear interaction. (b) Number of

runs in which a given fraction of the population spends less than control for eight-member household-

agents. Lines are as for Fig. 7(a).


losses were smaller than for the eight-member household-agents, being in the region

of 0.5� 0.4 tonnes.

In this model, four-member (better-off) household-agents do not starve. So, for

these household-agents, we looked at the yield rather than the cost. For any house-

hold class, only a small fraction of each population gets lower yields in a significant

number of runs. However, in up to one-third of runs, over 90% of the population

experience a net surplus. In this case the surplus can be significant, with a total mean

excess yield over the 50 years of up to 5� 0.5 tonnes. The effect of the forecast is inthis case felt more by the wealthier household classes, except that the four-member

household-agents have a peak in the yield improvement at 70%. For this group the

effect of non-linear interaction is to further improve yields (see Fig. 9(a) and (b)).

0

50

100

150

200

250

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Num

ber

of r

uns

Fraction of Population: 4 member households(a)

0

20

40

60

80

100

120

140

160

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Num

ber

of r

uns

Fraction of Population: 4 member households(b)

Fig. 9. (a) Number of runs in which a given fraction of the population has poorer yields than control for

four-member household-agents. Lines are as for Fig. 7(a). (b) Number of runs in which a given fraction of

the population exceeds yields of control for four-member household-agents. Lines are as for Fig. 7(a).


4. Discussion and suggestions for further work

4.1. Build up of trust

A somewhat surprising result in the above is the long time scale over which trust

builds up, despite the three-year rate at which farmer-agents learn to believe in the

forecast. The forecast accuracy is important, as can be seen from the following sim-

ple argument:Suppose that trust T increases linearly at rate i if the forecast is correct, and de-

creases linearly at rate d if the forecast is incorrect, and that the forecast is correct a

fraction p of the time. Then we can write for the mean trust hT i (over many se-

quences) after N years:


hT i ¼ N � ðp � i� ð1� pÞ � dÞ:

Clearly if p ¼ 1, then hT i ¼ N � i. This is the maximum possible, and we will reach a

mean hT i of 3 after three years if i ¼ 1. However, we believe that p is nearer 0.5 than

1, which gives

hT i ¼ 0:5 � N � ði� dÞ:
Now if i ¼ d, hT i ¼ 0 – nobody will (on average) trust a forecast that is only right
50% of the time. If p ¼ 0:6 and i ¼ d, then hT i ¼ 0:2 � N � i and now when i ¼ 1,

we only reach hT i of 3 after N ¼ 3=0:2 ¼ 15 years. This is still an optimistic estimate

of the mean if we assume that the trust level drops back to some small value when

those adopting the forecast get a bad result – the individual sequences will briefly

exceed T ¼ 3 and then fall below. This slows the rate of increase of hT i averaged overmany sequences still further.

Can we make the mean rate of adoption increase more rapidly? If i ¼ d and

p ¼ 0:5, the answer is clearly no – although the spread about the mean will be in-

creased. If p ¼ 0:6 and i ¼ d, then increasing i to 5 will get us to hT i ¼ 3 in 3 years

(ignoring drops for failed forecasts). However, this implies that individual people are

intrinsically ready to start using the forecast after only about 6 months rather than

the 3 years suggested by field studies. Alternatively, we might put d ¼ 0 (i.e., the limit

of i � d), so that people never reduce their trust when the forecast fails. Then withp ¼ 0:6 and i ¼ 1, we get hT i ¼ 3 after 5 years.

4.2. Suggestions for further work

There are a number of areas where the model described here could be extended.

First, there is a simplified representation of climate – rainfall is divided into three cat-

egories sampled with a crude frequency distribution, although such information is

typical of the level of detail available from consensus forecasts. Further, we take

no account of the effect of extremes that might lead to across the board crop failure

for all household classes over an entire country.

Second, the crop parametrization is highly simplified. To begin with, we have al-

lowed only for variation in rainfall, when temperature may have serious effects (un-expected frosts may destroy an entire crop). A more complete representation of crop

response to climate is clearly called for (see, e.g., Challinor et al., 2003, for an exam-

ple in the case of ground nut in India).

Third, although the response of people to the forecast itself and to each other has

been addressed in the model, it could be explored further. In the role-play a number

of response options were given and two of the most common ones were chosen for

use in the model. Given a crop model, we can say what the best response in a given

climate would be, but is this in fact what would be adopted on the ground? Farmersmay be reluctant to risk changes in farming practice in the face of untried new infor-

mation. Or, they may experiment with only part of their fields, reducing the costs of

poor forecasts. Moreover, the perception of the forecast must be set in the context of

local culture. Traditional beliefs about the climate and its representation in the local


environment may conflict with those suggested by the forecast. Although this was

only a problem for a small proportion of the role-play participants, cultural objec-

tion may be stronger in other regions. The presentation of the forecast by extension

officers, and the role of the village chief may either accelerate or hinder forecast

adoption. Further field work is needed to try to find ways in which these kinds ofeffects can be faithfully rendered by the modelling process. In the context of interac-

tion, more is needed in the exploration of opinion formation. It may be that al-

though only a few people are made worse off by a good forecast, they are vocal

or important community members. Such people may have a disproportionate effect

on forecast impact.

Fourth, the economy, either local or global, is hardly represented in the current

case. If people run out of food, are they able to buy it in? Is there sufficient local sup-

ply to allow this to happen? Many families own animals and gain income in otherways than through farming, either by share-cropping, selling whatever they can

make from local materials, or else by seeking work outside the village. All of these

factors not only determine how vulnerable a household may be to crop failure,

but will affect their attitude to and ability to take on risk, and thus their ability to

deal with the possible benefits of the forecast.

Last, finer time steps of farmer-agent responses could be included. For example,

monthly timesteps could be incorporated to enable the initial strategy to be updated

depending on how the predicted and actual climate compare through the season.This could enable disseminators to assess what impacts an improvement on the tim-

ing of dissemination would have, something users are concerned with.

5. Conclusions

The paper set out to achieve three aims. The first aim was to evaluate the pay-offs

of using the forecast as a function of forecast accuracy. The model shows that unlessthe forecast accuracy is 60–70% or above, positive impacts from using forecasts are

unlikely. Forecast accuracy is therefore paramount in determining the direct benefit

of forecast use to marginal farmers. Additional to this, the model suggests that non-

linearities in crop response (in this case to rainfall) are crucial to the usability of fore-

casts.

The second aim was to establish how the mean trust of individual farmer-agents

changes as a function of forecast accuracy. The results show that the level of forecast

accuracy determines the level of trust and that there is a threshold of below whichtrust will be lost significantly. If the forecast is accurate 60% of the time and very

poor not more than 10% of the time, then trust saturates at 30 years. For some

household-agents, the mean trust level saturates at the point where using the forecast

begins to become economically advantageous. Many household-agents, if acting in

isolation, may give up using forecasts earlier because of short-term losses, even

though in the longer term using forecasts may be to their benefit.

The third aim was to measure the benefit of using the forecast in terms of yield,

with and without farmer-farmer interaction and for different wealth profiles. Results


indicate that social processes have an impact on the effect of the forecast, both from

the point of view of individual response and that of interaction between individuals.

Poor farmer-agents benefit the most in the sense of reducing the likelihood of food

shortage conditions (cf. Figs. 5 and 8(a) and (b)), but yield enhancement is better for

the wealthier household classes.The ABSS model highlights some of the potential impacts that seasonal forecasts

might have on rural households in Lesotho. Since the model explores possible out-

comes of individuals using information that they do not presently use, the results

are hard to validate. However, field work and theoretical evidence supports many

of the results. It suggests that seasonal climate forecasts are not a panacea to improv-

ing poor households’ rural livelihoods.

Although the model is a pilot and based on data from one area in Lesotho, it pro-

vides a holistic approach to assessing the impact of anticipatory information. Themodel is not being used for policy at this stage, so its merit is more in the exploratory

realm of investigating possible ramifications of using a forecast of varying skill in dif-

ferent ways. Policies for sustainability cannot wait until all facts are known. Rather,

we need to use information that is as robust as possible to plan and implement

changes in technology and activities. By explicitly documenting the steps and asso-

ciated assumptions used to produce the model, as well as listing the simplifications,

we hope that it can be re-evaluated for another case, or if different factors are a pri-

ority that these can be changed at the appropriate level.The evolution of trust highlights the importance of understanding the human

agency of users and how it varies. Opinion formation and the establishment of trust

is complex and forecast disseminators cannot expect to have credibility if they do not

explain the probabilistic nature of the forecast and past forecast skill. The results

also imply that a certain level of forecast skill is required in order for trust not to

decline steadily. Forecast centres and disseminators might consider improving fore-

cast skill before disseminating seasonal forecasts. They may also need to recognise

that large forecast errors have differential impacts on trust. If the climate predictionis bi-modal with low confidence, forecast centres may wish to make sure users are

aware of the uncertainties. If trust is lost, it may take some time be regained.

The model draws attention to the fact that villages cannot be treated as homog-

enous. The impact of the forecast on better-off and poor households differs signifi-

cantly. If national level forecast dissemination strategies are prioritised, the model

suggests that blanket targeting of rural groups will not necessarily address user

needs. Further investigation into the impact of the forecast at the village level is

encouraged before dissemination is actively pursued without appreciating potentialimpacts.

Acknowledgements

G.Z. thanks the School of Geography and the Environment, Environmental

Change Institute and the Rhodes Trust, University of Oxford for funding this re-

search. The contributions from the people of Ha Tlhaku and CARE Lesotho are


gratefully acknowledged. Matthew Swann and Edmund Chattoe are thanked for

their useful comments and help. The model development is continuing in a case

study in South Africa, funded by the Tyndall Centre (see www.geog.ox.ac.uk/re-

search/projects/cloud) and those contributors are thanked for their support to this

work. The conclusions reached are solely those of the authors.

References

Agrawala, S., Broad, K., Guston, D.H., 2001. Integrating climate forecasts and societal decision making:

challenges to an emergent boundary organisation. Science Technology and Human Values 26, 454–

477.

Anderson, J.R., 2003. Risk in rural development: challenges for managers and policy makers. Agricultural

Systems 75 (2–3), 161–197.

Archer, E.R.M., 2003. Identifying underserved end-user groups in the provision of climate information.

Bulletin of the American Meteorological Society 84 (11), 1525–1532.

Barreteau, O., Bousquet, F., Attonaty, J.-M., 2001. Role-playing games for opening the black box of

multi-agent systems: method and lessons of its application to Senegal River Valley irrigated systems.

Journal of Artificial Societies and Social Simulation 4.

Basher, R., Clark, C., Dilley, M., Harrison, M., 2001. Coping with Climate: A Way Forward.

International Research Institute for Climate Prediction, Palisades.

Blench, R., 1999. Seasonal climatic forecasting: who can use it and how should it be disseminated? Natural

Resource Perspectives 47 001, London.

Bureau of Statistics, and Food and Agriculture Organisation, 2001. Lesotho Agricultural Census 1999/

2000: Rural: Households and Crop Statistics, vol. I. Maseru.

Cane, M.A., Zebiak, S.E., Dolan, S.C., 1986. Experimental forecast of El Ni~no. Nature 321, 827–823.

Carberry, P., Hammer, G.L., Meinke, H., Bange, M., 2000. The potential value of seasonal climate

forecasting in managing cropping systems. In: Hammer, G., Nicholls, N., Mitchell, C. (Eds.),

Applications of Seasonal Climate Forecasting in Agricultural and Natural Ecosystems – The

Australian Experience. Kluwer Academic Publishers, Dordrecht.

Chakela, Q.K. (Ed.), 1999. State of Environment in Lesotho 1997. National Environment Secretariat,

Maseru.

Challinor, A.J., Slingo, J.M., Wheeler, T.R., Craufurd, P.Q., Grimes, D.I.F., 2003. Towards a combined

seasonal weather and crop productivity forecasting system: determination of the working spatial scale.

Journal of Applied Meteorology 42 (2), 175–192.

Chattoe, E., 2000. Why is building multi-agent models of social systems so difficult? A case study of

innovation diffusion. In: XXIV International Conference of Agricultural Economists (IAAE),

Mini-Symposium on ‘‘Integrating Approaches for Natural Resource Management and Policy

Analysis: Bioeconomic Models, Multi-Agent Systems and Cellular Automata’’, Berlin, August 13–

19.

Downing, T.E., Moss, S., Pahl-Wostl, C., 2000. Understanding climate policy using participatory agent-

based social simulation. In: Moss, S., Davidsson, P. (Eds.), Multi-Agent-Based Simulation: Second

International Workshop. Springer, Boston, MA.

Eakin, H., 1999. Seasonal climate forecasting and the relevance of local knowledge. Physical Geography

20, 447–460.

Eakin, H., 2000. Smallholder maize production and climatic risk: a case study from Mexico. Climatic

Change 45, 19–35.

Edmonds, B., Moss, S., Wallis, S., 1996. Logic, reasoning and a programming language for simulating

economic and business processes with artificially intelligent agents. CPM Report No.: 96-09, Centre for

Policy Modelling, Manchester.

Gadgil, S., Seshagiri Rao, P.R., Narahari Rao, K., 2002. Use of climate information for farm-level

decision-making: rainfed groundnut in southern India. Agricultural Systems 74 (3), 431–457.

http://www.geog.ox.ac.uk/research/projects/cloud

http://www.geog.ox.ac.uk/research/projects/cloud


Gay, J., Hall, D., 2000. Poverty and livelihoods in Lesotho, 2000. More than a mapping exercise. Sechaba

Consultants, Maseru.

Hansen, J.W., 2002. Applying seasonal climate prediction to agricultural production (preface).

Agricultural Systems 74 (3), 305–307.

Hegselmann, R., Krause, U., 2002. Opinion dynamics and bounded confidence: models, analysis and

simulation. Journals of Artificial Societies and Social Simulation 5. Available from <http://

jasss.soc.surrey.ac.uk/5/3/2.html>.

Ingram, K.T., Roncoli, M.C., Kirshen, P.H., 2002. Opportunities and constraints for farmers of west

Africa to use seasonal precipitation forecasts with Burkina Faso as a case study. Agricultural Systems

74 (3), 331–349.

Kirshen, P., Flitcroft, I., 2000. Use of seasonal precipitation forecasting to improve agricultural

production in the Sudano-Sahel: an institutional analysis of Burkina Faso. Natural Resources Forum

24, 185–195.

Martin, R.V., Washington, R., Downing, T.E., 2000. Seasonal maize forecasting for South Africa and

Zimbabwe derived from an agroclimatological model. Journal of Applied Meteorology 39,

1473–1479.

Mason, S.J., Joubert, A.M., Cosijn, C., Crimp, S.J., 1996. Review of seasonal forecasting techniques and

their applicability to southern Africa. Water SA 22, 203–209.

Mazzocco, M.A., Mjelde, J.W., Sonka, S.T., Lamb, P.J., Hollinger, S.E., 1992. Using hierarchical systems

aggregation to model the value of information in agricultural systems: an application for climate

information. Agricultural Systems 40, 393–412.

Miller, C., 1998. Extending assessment communities to developing countries. John F. Kennedy School of

Government, Harvard University.

Moss, S., 1999. Relevance, realism and rigour: a third way for social and economic research. CPM Report

No.: 99-56, Centre for Policy Modelling, Manchester.

Mukhala, E., Walker, S., van den Berg, J., 2000. Communication of seasonal climate forecasts between

meteorological scientists and farmers in the Free State Province. In: Southern African Climate Outlook

Forum. Gaborone, Botswana.

Murphy, A., 1997. Forecast verification. In: Katz, R., Murphy, A. (Eds.), Economic Value of Weather and

Climate Forecasts. Cambridge University Press, Cambridge, UK, pp. 19–74.

Murphy, S., Washington, R., Downing, T.E., Martin, R.V., Ziervogel, G., Preston, P., Todd, M.,

Butterfield, R., Briden, J., 2001. Seasonal forecasting for climate hazards: prospects and responses.

Natural Hazards 23, 171–196.

Nicholls, N., 1999. Cognitive illusions, heuristics, and climate prediction. Bulletin of the American

Meteorological Society 80, 1385–1397.

O’Brien, K.L., Sygna, L., N�ss, L.O., Kingamkono, R., Hochobeb, B., 2000. Is information enough? User

responses to seasonal climate forecasts in Southern Africa. Report No. 2000–03, CICERO.

Orlove, B., Tosteson, J., 1999. The application of seasonal to interannual climate forecasts based on

ENSO events: lessons from Australia, Brazil, Ethiopia, Peru, and Zimbabwe. Working Paper 99-3.

Institute of International studies, University of California, Berkley.

Palmer, T.N., Anderson, D., 1994. The prospects for seasonal forecasting – a review paper. Quarterly

Journal of the Meteorological Society 120, 755–793.

Patt, A., 2000. Communicating probabilistic forecasts to decision makers: a case study of Zimbabwe.

Report No. 2000-19, Belfer Center for Science and International Affairs (BCSIA), Environment

and Natural Resources Program, Kennedy School of Government, Harvard University,

Cambridge.

Patt, A., 2001. Understanding uncertainty: forecasting seasonal climate for farmers in Zimbabwe. Risk,

Decision and Policy 6, 105–119.

Patt, A., Gwata, C., 2002. Effective seasonal climate forecast applications: examining constraints for

subsistence farmers in Zimbabwe. Global Environmental Change: Human and Policy Dimensions 12,

185–195.

Peshoane, L.A., 2000. Seasonal rainfall forecast for OND 2001/JFM 2002. In: Lesotho Meteorological

Services National Seasonal Forecast Workshop, Maseru.

http://jasss.soc.surrey.ac.uk/5/3/2.html

http://jasss.soc.surrey.ac.uk/5/3/2.html


Phillips, J.G., Makaudze, E., Unganai, L., 2001. Current and potential use of climate forecasts for

resource-poor farmers in Zimbabwe. In: Impacts of El Nino and Climate Variability on Agriculture,

vol. 63, American Society of Agronomy Special Publication Series, pp. 87–100.

Rogers, E.M., 1995. Diffusion of Innovations, fourth ed. The Free Press, New York, NY.

Rowland, J.R.J. (Ed.), 1993. Dryland Farming in Africa. Macmillan Education Ltd., London.

Smit, B., Burton, I., Klein, R.J.T., Wandel, J., 2000. An anatomy of adaptation to climate change and

variability. Climatic Change 45, 223–251.

Stern, P., Easterling, W. (Eds.), 1999. Making Climate Forecasts Matter. National Academy Press,

Washington, DC.

Stone, R.C., de Hoedt, G.C., 2000. The development and delivery of current seasonal climate forecasting

capabilities in Australia. In: Hammer, G.L., Nicholls, N., Mitchell, C. (Eds.), Applications of Seasonal

Climate Forecasting in Agricultural and Natural Ecosystems – The Australian Experience. Kluwer

Academic, The Netherlands, pp. 67–75.

Turner, S., Calder, R., Gay, J., Hall, D., Iredale, J., Mbizule, C., Mohatla, M., 2001. Livelihoods in

Lesotho. CARE Lesotho, Maseru.

Vogel, C., 2000. Usable science: an assessment of long-term seasonal forecasts among farmers in rural

areas of South Africa. South African Geographical Journal 82, 107–116.

Walker, S., Mukhala, E., van den Berg, J., Manley, C., 2001. Assessment of communication and use of

climate outlooks and development of scenarios to promote food security in the Free State Province of

South Africa. Department of Agrometeorology, University of the Free State, Bloemfontein.

Washington, R., Downing, T.E., 1999. Seasonal forecasting of African rainfall: prediction, responses and

household food security. The Geographical Journal 165, 255–274.

Whiteman, P.T.S., 1981. Sorghum and millet agronomy investigations in Eastern Province: results and

discussion of trials. Technical Report KEN/78/016, FAO, Katumani, Kenya.

Ziervogel, G., 2004. Targeting seasonal climate forecasts for integration into household level decisions: the

case of smallholder farmers in Lesotho. The Geographical Journal 170, 1.

Ziervogel, G., Calder, R., 2003. Climate variability and rural livelihoods: assessing the impact of seasonal

climate forecasts. Area 35 (4), 403–417.

Ziervogel, G., Downing, T.E., in press. Stakeholder networks: improving seasonal forecasts. Climatic

Change.

Documents

Agent-based social simulation: a method for assessing the ...web.csag.uct.ac.za/~gina/Gina_Ziervogels...Agent-based social simulation: a method for assessing the impact of seasonal