Forecasting Public-Project Expenditure with Limited Historical Data: A Reproducible Simulation Study

Financial forecasting is often constrained by short histories, variable spending cycles, and sudden changes in planned activity. This methodological study compares seasonal-naive forecasting, regularised linear regression, and random forests in three fully synthetic monthly-expenditure settings: regular seasonal expenditure, an unexpected level shift, and high volatility Six replicates per setting generate 120 observations each. Predictions use a rolling expanding training window and 24 one-month-ahead test periods per replicate, yielding 432 forecast-target observations and 1,296 method-specific predictions.Mean absolute error is evaluated in simulation units. Across the combined scenarios, ridge regression obtains a mean absolute error of 19.64, versus 24.15 for random forests and 29.40 for seasonal-naive predictions. The results illustrate that a modest, interpretable model can outperform a more flexible model when historical data are limited and trends matter. Because the data are synthetic and the design is narrow, the ranking should not be generalised to Pakistani budget systems without real-data validation.