Forecasting Daily January Temperatures in London Using Nine Statistical Approaches

Precise prediction of winter temperature is crucial for public safety and urban energy management, especially under the effects of global warming. This article compares nine different time series forecasting models, trained on historical data from 1979 to the end of 2019, to predict London's January mean temperature in 2020. The predicting approach includes the historical and filtered mean method, drift method, regression-based method, ARIMA-family models, and non-linear methods. Performance was evaluated via MAE and RMSE against the actual observation. According to these two metrics, Two-Stage Regression, SARIMA with Typical Day Reconstruction, and Filtered Mean are the three most accurate models for this London weather dataset. Residual diagnostics, including the LjungBox test, indicate significant autocorrelation despite the ACF plot resembling a random walk, reflecting model limitations. To ensure robustness, expanding- and rolling-window cross-validation was conducted on the three best-performing models, alongside DieboldMariano and paired t-tests for statistical comparison. Finally, a linear trend analysis on the January mean temperature was employed to identify a long-term climate pattern. This research shows and compares a range of forecasting methods, which offer perspectives into real-world applications such as climate adaptation.