Millised on probleemid ajalooliste mudelite tagasivaatamisel?

During backtesting of time series models, the following problems arise:

• Data leakage: using future data to train the model, which is unacceptable for time series as it violates the chronology.
• Instability of time series: seasonality, trends, and structural shifts can affect the quality of the model and its ability to generalize.
• Choosing the correct training and testing window: it is necessary to properly split the data over time to avoid overlaps and ensure a fair assessment.
• Autocorrelation of errors: model errors can be dependent over time, complicating the evaluation of quality.
• Change in data distribution (concept drift): the behavior of the time series can change over time, and a model trained on old data may perform poorly on new data.

For proper backtesting, methods such as rolling window or expanding window are often used to simulate the actual forecasting process.