Forecasting in e-Wallet Industry

Our Customer who’s a leading Fintech company came up with a use case of Forecasting for their user base. That would get converted from 1 Biller to 2 and then 3 and so on. They wanted us to assist them in strategizing their biller strategy. 

In order to achieve the forecasting of users there was a need to first forecast the Biller counts across the board. Hence the team at Aspire NXT helped Fintech to Forecast the biller counts at a   better and higher accuracy using Amazon Forecast. 

Fintech Forecast Use case

  • Given a timeseries data containing, timestamp, bill category and  biller count per category; The goal is to forecast the biller count per category over a period of time. 
  • In retrospect, build and train a forecasting model using Amazon Forecast to achieve the same.

Amazon Forecast

  • Amazon forecast is an accurate time-series forecasting service, based on the same technology used at Amazon.com 
  • Forecasting at Amazon provides the solution using machine learning to solve hard forecasting problem 
  • With Amazon Forecast, you can achieve forecasting accuracy levels that used to take months of engineering in as little as a few hours.

Amazon Forecast Advantages 

  • Automated Machine Learning Service for accurate forecasting.
Fully Managed Service

Automatically set data pipeline, training and prediction

Highly Accurate

50% increase in accuracy over traditional methods

Easy to Use

No deep learning experience required

Highly Secured

Encrypted through Amazon KMS

Forecasting in E-Wallet Industry

  • Sales Forecasting – predict sales over a period and boost margins.
  • Increase in Biller Countforecasting per billing category of customer base.
  • Forecast Industry 4.0 – provide detailed analysis of the market structure along with forecast of the various segments and sub-segments of the Industry 4.0 market.
  • Forecast service usage bill and customer engagement.
  • Resource planning for the optimization of available resources, such as staffing levels, advertising inventory to maximize revenue and control costs
  • Product cost reduction and ROI increase.

Architecture

Fintech Data Exploration

  • Data was feature engineered on Redshift cluster to a format expected by Redshift and was unloaded to S3.
  • Exploratory data analysis was carried out on Sagemaker, as shown in the attached visuals.
  • The target time series data and a validation file is prepared.

Fintech Data Exploration

  • After preliminary EDA, three predictors were trained on different algorithms (ARIMA, Prophet & DeepAR+) provided out of the box in Amazon Forecast.
  • The model was trained on the parameters as shown below
  • The RMSE values obtained for each of the models are as shown below. It is clear that  DeepAr+ has the least RMSE and is consistent over all the test windows.

Weighted Quantile Loss

  • Amazon Forecast provides probabilistic predictions at three distinct quantiles by default
    —10%, 50%, and 90%—and calculates the associated loss (error) at each quantile.
  • The weighted quantile loss(wQuantileLoss) calculates how far off the forecast is from actual demand in either direction.
  • From 10thand 90th percentiles we can be sure that 80% of the time the actual value will be within this boundary range. If need be, this accuracy can be further increased.
  • The quantile losses at p50 are as shown:
  • From these p50 quantile losses and RMSE values it can be concluded that DeepAR+ performed the best and is an optimal algorithm for this use case.

Fintech Forecast Results

  • DeepAR+ Forecast

Fintech Forecast Results

Other Bill Payment Bill Count

Telco Bill Payment Bill Count

Zakat donations Bill Count

  • Surges in data lie below the P90 quantile as per expectation however, errors in prediction is also visible and this can be surmounted by training the predictor with more data points.
  • Zakat on the other hand needs significant data points for the predictors to gain insight on the trend

Arima Validation

Other Bill Payment Bill Count

Telco Bill Payment Bill Count

Zakat donations Bill Count

  • ARIMA algorithm performs best as the target value generally coincides with the p50 quantile
  • Zakat on the other hand needs more data points to be trained

Prophet Validation

Other Bill Payment Bill Count

Telco Bill Payment Bill Count

Zakat donations Bill Count

  • Prophet algorithm performance is similar to that of ARIMA however all 3 algorithms need more data points for Zakat donations category.

Comparison & Conclusions

  • For  smaller size of data, ARIMA follows the trend better, but cannot predict the spikes well
  • DeepAR+ gives a closer prediction to the spikes.
    • With more data and perhaps related data series, this behaviour will be improved.
    • Trend in datais followed in general for Other and Telco bill payments
  • Furthermore, quantity of data also needs to be increased in order to obtain better prediction results.
    • Expected data pointsfor the optimal performance of the predictor is 1000 while this use case had 360 unique data points

Comparison & Conclusions

  • Given that the quantity of the data for this POC was small (approximately 10KB) the cost incurred was 0.29USD based on 1.204hrs of predictor training
Top