I have performed Boxcox transformation on my time series data and processed it through ARIMA modeling. Converted prediction values to the actual. I see significant differences between actual and predictions(In sample), would like to get optimal prediction values which are close to actual values.
Have tried Machine learning models such as Linear Regression, tree-based(Random Forest, XGboost) as well. Before processing data through these models, it(sales) was aggregated by day wise and extracted, sale day, month, sale dow..etc.
For the same data, lagged features and rolling statistics were also generated, trained and tested. However, did not get optimal sale predictions and found differences between actual sale values and predictions.
I have ensure no white noise in my data and made stationary before processing data thru models.
Can someone suggest techniques such as feature engineering or any other modeling to overcome the problem of differences between actual and predictions and future forecasting?
Actual Predcition 72 310.255 2148 1246.419 208 497.629 531 257.541 80 125.156 480 464.411 229 365.513 60 296.102 192 279.578 90 278.131 72 180.8 70 48.544 16 341.464 32 225.55 1560 253.01 960 708.296 1528 498.59