ENHANCED LSTM MODELS FOR SHORT-HORIZON FORECASTING: FROM FINANCIAL TIME SERIES TO PUBLIC GOVERNANCE

Povzetek

This research paper evaluates the performance of an improved Long Short-Term Memory (LSTM) model in forecasting stock prices for Google and Microsoft over varying timeframes (1-year, 2-year, and 3-year). The improved LSTM model consistently outperforms the original model for Google, demonstrating significant reductions in Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) across all timeframes. In contrast, the results for Microsoft are mixed; the improved model exhibits substantial performance gains for the 1-year data but underperforms relative to the original model for the 2- year and 3-year periods. This discrepancy suggests potential overfitting or limitations of the improved model in capturing long- term trends for Microsoft. The findings highlight the importance of model adaptability when applied to different stocks and time horizons, offering insights into the strengths and challenges of using advanced LSTM architectures in financial forecasting. This paper also further demonstrates how the improved LSTM pipeline can be adapted to public-sector short-horizon forecasting problems (e.g., municipal property-tax receipts, short-term water demand, daily waste generation, and clinic throughput). We describe domain adaptations — such as domain-specific lag windows, GIS/spatial features, exogenous policy and weather indicators, and uncertainty quantification (prediction intervals) — that are required for responsible deployment in local self-government contexts. These extensions show the technique’s operational value for municipal budgeting, resource allocation, and short-term operational decision-making, and motivate pilot deployments with co-designed evaluation metrics oriented to administrative use.