New Techniques to Perform Cross-Validation for Time Series Models

Model validation for time series models has always been a challenge due to a lot of complexities. The presence of auto-correlation in the data creates a challenge to the conventional cross validation techniques like k-fold cross validation to be implemented for time-series models. In this paper, two weighted k-fold time series split cross-validation techniques are proposed for this purpose. The proposed techniques were validated using the opening price data of cryptocurrency. Mean squared error (MSE), Mean absolute error (MAE) and Mean absolute percentage error (MAPE) were the selected metrics to validate the proposed techniques. Both the techniques were found to give robust results; however, the Exponential weighted K-fold time series split cross validation (EWKCV) technique was seen to perform better than Generally weighted K-fold time series split cross validation (GWKCV) technique. The results of the proposed techniques, along with the results of simple train-test split for the time-series models, is seen to give better result.