earticle

영어

Time series forecasting is relevant in many real-world applications. However, most real-world time series data are non-stationary, which means their statistical properties like mean, and variance varies with time. This property of time series is not considered by most modern deep learning forecasting models, causing the distribution of the training and test sets to be different. Eventually, the accuracy of the forecasting model is significantly affected by the distribution shift. To tackle this problem, we suggest a simple solution called 'Pseudo-Stationarizer.’ This block can be used seamlessly alongside pre-existing forecasting models to obtain better forecasts. ‘Pseudo-Stationarizer’ performs differencing on the original time series to make the data weakly stationary and helps in minimizing the distribution shift. Via thorough experimentation, we prove that the usage of the proposed block aids the forecasting models in getting significant improvements in their performance by diminishing the distribution shift and making the time series weakly stationary.