Upload
others
View
3
Download
0
Embed Size (px)
Citation preview
Group 7104078515 Chiali 104078509 Sherry
105078514 Elisa 105078502 Emily
Forecasting Daily Number of User Problem Reports of Junyi Academy for Efficient Staff Allocation
2017.1.3
Background
Junyi Academy is a platform offering online learning resource for all ages! It provides practice exercises, instructional videos, and a personalized learning dashboard that empower learners to study at their own pace in and outside of the classroom.
With high utilization ratio of the practice exercises, Junyi receives problems reported by users, which is called “user problem reports”. And all the reports will be checked and then be distributed to the responsible team by operation team.
Since there is no full-time staff dealing with user problem reports. We want to help them better allocate manpower, so that every report could be resolved on time.
We are forecasting (Ft): the daily number of user problem report in the coming week
Yt: daily number of problem reportForecast horizon: h=1,2,3,4,5,6,7 (a-week-ahead forecasts)Goal type: forward-looking
Business & Forecasting Goal
With the forecast of the number of user problem reports, the manager could better allocate staff and their working loading.
Client: The manager of Junyi Academy Operation TeamStakeholder: Junyi’s employees, Students, Teachers, Parents
BusinessGoal
Forecasting Goal
There is no full-time staff dealing with user problem reports. Thus, if daily reports are over 23 , it is very likely that reports cannot be solved on that day.(which is the average
number of reports solved)
Business & Forecasting GoalData
Source: Junyi AcademyTime Period: 2016.08.29~2016.11.13Amounts of Rows: 77Series:
● Received: daily number of problem report
(with weekly seasonality and no trend)
● New registration user: daily number of new registered user
● Active user: daily number of active user, who finished one exercise on that day
Business & Forecasting GoalData Preprocessing
● Create seasonal dummies
● Create binomial series for school-day
● Create lag series: lag7_Registration and lag7_ActiveUser
● Find the outlier and create a binomial series for outlier
- Based on boxplot of series “Received”
- Typhoon day-off
Methods
MLR & NN have better and similar performance
Select methods (based on the performance on RMSE and MAPE)
SES, MA, Holt-Winter’s (no trend), MLR, NN
Try forecasting methods which can deal with series with seasonality but no trend
Seasonal Naive Benchmark
Methods & Evaluation
T r a i n i n g V a l i d a t i o n
RMSE MAPE RMSE MAPE
Seasonal Naive [ Benchmark ] 8.8784 30.3077 6.4031 28.7795
SES 9.4985 33.1938 7.3374 28.3997
Trailing MA 10.4972 35.7355 9.1741 31.5272
Holt-Winter (ANA) 7.0517 24.3907 6.2843 25.4447
Regression Model with external data 5.4104 21.8819 5.3927 21.5305
Neural Network (5-25-1) 3.4786 13.3547 5.1661 20.9879
Methods
MLR & NN have better and similar performance
Select methods (based on the performance on RMSE and MAPE)
SES, MA, Holt-Winter’s (no trend), MLR, NN
Try forecasting methods which can deal with series with seasonality but no trend
Seasonal Naive Benchmark
Try different modelsdifferent combination of external series
MLR
Try different models with different combination of external series and deseasonalizing
Not include lag-7 Registrationpredictor: Lag-7_Active user, SchoolDay, Outlier,
Seasonal dummies
Include all external data and seasonal dummies
NN
Based on the time plot and ability of capturing peaks, we chose this model.
Methods & EvaluationMLR with external series
NN1 hidden layers & 25 neuronsusing XLMiner v.s. R
Results
NN - Training
RMSE3.5912
MAPE13.7274
NN 1 hidden layers & 25 neurons using R
23
Limitations & Recommendations
Limitations● We cannot forecast the exact number of peak
values, but we can capture where the peaks are.
● There will be lower cost for over-forecast than under-forecast. However, our model tends to be under-forecast at the peaks.
○ Over-forecast: Extra part-time workers hired to deal with reports will be assigned with other tasks.
○ Under-forecast: Staff will not be able to finish their work on time and left problem reports unsolved.
RecommendationsAccording to the limitations, we suggest Junyi could run our predictive model on every Friday to forecast the next week’s report number.
Thus, it will help them● decide whether to allocate extra staff
to deal with reports● arrange people to check the content
of the questions before they are released.