Benchmarking Machine-Learning Models for Weekly River-Sand Price Floor Prediction

Abstract

River-sand price volatility in Sri Lanka forces sub-optimal procurement and erodes margins. We benchmark seven machine-learning families for next-week floor-price forecasting across six depots and test whether sparse linear regularization outperforms non-linear alternatives under small-sample conditions. We assemble the first VAT- and transport-neutral weekly panel of river-sand quotes from six Kalutara depots (3 June 2018–23 February 2020), sourced from invoices, quotations, and ledger scans and harmonized into a continuous series. A feature-rich time-series pipeline creates lagged minima (1–4 weeks), rolling minima (7, 14, 30 days), and harmonic calendar encodings. Seven algorithms—ElasticNet, Ridge, Random Forest, HistGradientBoosting, Gradient Boosting, K-Nearest Neighbors, and Support Vector Regression—plus a naïve persistence baseline are evaluated using expanding-window cross-validation to mirror real-world information flow. Within this benchmark, ElasticNet yields the lowest hold-out MAE (LKR 4.50; 99.2% below the baseline MAE of LKR 563) and R² ≈ 0.99996, relying on just three predictors: the previous week’s minimum, the 14-day rolling low, and a post-budget flag. This meets pre-stated parsimony and stability criteria for deployment at scale. Non-linear ensembles and instance-based learners underperform, often failing to beat the trivial baseline, highlighting small-n overfitting risk in frontier markets. For practitioners, a simple “buy when spot ≤ forecast + 0.2%” rule converts model outputs into average savings of LKR 600–800 per cube in historical back-tests. Overall, aligning model complexity with data reality delivers interpretable forecasts and tangible cost savings in practice.