We are trying to predict customer’s Total recharge value for the next week. We came up with 8 recharge bands. So its now a classification problem. We have tried history of last 8 weeks(Weekly aggregated attributes). We have tried all the famous techniques like feature selection, PCA, feature engineering, tried different algorithms, balancing, boosting, bagging. But the overall accuracy is always below 45%. How can we improve it ? Also can time series be of any help for this problem?
if u have trnxs data, i think you can also adopt RFM approach to predict total value recharge by users in one week time period. there is a very good package in R called BTYD, it can help you predict the value of recharge a customer will go for in specific time period (day, week, month, year). you got to have user_id, trnx_date, amount variables in data & go for BTYD. you will also be able to get churn, LTV etc out of it
Thanks for the help but we are using SPSS Modeler without the integration with R.
please is there any update of this question?
I tried BTYD package in R but it did not perform well !!
you can use ensemble method.That is combining various models to improve the accuracy.
Also validation is important for the sample you take.try various sampling techniques like kfold sampling etc.