Bank Business Volume Time Series Forecast

Introduction

In order to efficently allocation the sources of bank, it is very important for bank to predict the future business volumes.The task of this competition is to predict the future business volumes from day granularity(sub task 1) & hour granularity(sub task 2).

Data

the overall data structure is simple,just in form of date -- period in the day -- volume.

There is another table to tell us the type of each day, for example, workday, weekend, hoilday and so on.

What needs additional attention is that there are two types of businesses,A and B,we need to deal with them respectively.

The training data includes the data from 2018.1.1 to 2020.10.31, and our prediction targets are the business volumes for the day between 2020.11.1 to 2020.12.31.

Basic Solution

1.Preprocess and EDA

Load data from csv and make some plot to observe the overall tendencies

2.Feature engineering

Besides the basic features the data provides like the date(year,month,day,hour,day type etc.),I also try to create some other features I think may help.

For example:

some dummies:

i.whether the previous day is weekend

ii.whether the previous day is hoilday

information of the previous days:

iii.the proportion of hoildays of the last 7 days

iv.the proportion of workday-to-rest of the last 14 days

and so on.

3. Modeling

I try to use 3 types of modeling methods.

1.The first is traditional time series forecast method like ARIMA, unfortunately, the performance of the model on validation set is not good.

2.The second is deep learning method,LSTM.The performance of it on validation set is also not good,the possible reason is the data size is not large enough to fully give play to the power of deep learining.

3.The final method is traditional machine learning method, I use 3 different ensemble learning methods.1)Random Forest 2)GBDT 3)Xgboost

Together with the features I made before, these 3 models' performance on validation set far beyond the traditional time series method and deep learning methods.

Finally,I use the combination of these 3 models to predict the data on test set. And my final score(MAPE) on task 1 & 2 is:0.136 & 0.7913 respectively.

4.Performance on validation set

Task 1 where the red line means the real values and the blue line means the predicted values.

Random Forest

GBDT

Xgboost

Task2

Xgboost