MODELING RETAIL SALES USING AUTOREGRESSIVE INTEGRATED MOVING AVERAGE AND LONG SHORT-TERM MEMORY FORECASTING METHODS

Background . Forecasting retail sales is crucial for modern supply chain and inventory management. Traditional statistical models alone can be insufficient due to the large amounts of data generated by extensive retail chains. Combining time series analysis with machine learning can improve forecast accuracy. Materials and Methods. This research used the M5-forecasting accuracy dataset, containing over 30,000 time series of store-item daily sales. The study involved data preprocessing to handle any missing values and splitting the series into training and hold-out test sets. Three forecasting methods were applied. The first method accounted for autoregressive and moving average components. The second approach explicitly included trend and seasonality by decomposing the series into those components, fitting a model to the trend-adjusted series, and then reintroducing the seasonal part. Third, a long short-term memory deep learning regressor was trained to capture longer-range dependencies. The evaluation on the test set was performed using the Mean Absolute Error (MAE). Residual analysis examined autocorrelation and the distribution of errors. Results and Discussion. A focus on one item showed a strong weekly cycle. The first autoregressive approach without explicit seasonality partially captured the data but left some significant autocorrelation in the residuals. The second autoregressive variant that considered trend and weekly seasonal decomposition achieved the best short-term predictive accuracy, reflected by lower MAE. The deep learning regressor, implemented in a recursive multi-step setup, did not outperform the autoregressive one, partly due to error accumulation and possibly incorrect choice of its architecture. Conclusion. The study indicates that for retail data with clear weekly fluctuations, auto­regressive moving-average models enhanced by trend and seasonal decomposition can provide robust forecasts. Neural network methods can model non-linearities but require mo­re specialized sequence-to-sequence configurations to avoid cumulative forecast errors. Fu­ture work can involve combining methods for multi-horizon and hierarchical retail time series. Keywords : Time series analysis, machine learning, retail forecasting, ARIMA, LSTM, seasonality