Show what a market forecast depends on

Early work focused on testing linear forecasting models against historical agricultural data. It combined market, weather, production and satellite features, with explicit comparisons against simpler benchmarks. The resulting research note kept the linear models as benchmarks rather than recommending them as the final production approach.

The next question was more explanatory: could the model estimate the supply and demand conditions behind a price, rather than only fit a price series? PitchAI developed a separate factor-based pipeline to investigate that question.

The pipeline forecasts processing demand, the contract-versus-free-market demand split, free stock, harvest volume and the share of usable quality. Those factors feed analytical measures of market tightness, which are then mapped to price pressure.

This creates intermediate results that can be examined. A poor price estimate can be investigated through the stock forecast, the quality estimate or the final mapping, instead of treating the whole model as one unexplained prediction.

The chained evaluation uses rolling prediction dates. Factor inputs are restricted by their dates, while the price estimator fits earlier seasons and evaluates later ones. Current and previous prices are excluded from the five-factor forecast inputs; historical prices remain necessary for fitting and evaluating the final mapping.

The code audits forbidden inputs, future target factors and selected source-availability violations, and raises an error when those checks fail. The retained run includes forecast inputs, predictions, error breakdowns and a separate provenance audit. These records make the experiment inspectable without relying on a chart alone.

Evaluation reports separate forecast horizons, points in the storage season and individual seasons. In the reviewed historical run, an aggregate two-month result passed its research threshold, but several narrower slices failed. Its product-readiness verdict therefore remained negative.

The additional value is the research system around the model: explicit intermediate factors, dated inputs, retained predictions and checks that can overturn a promising headline. It gives further development a concrete set of failure cases to address.

This is implemented forecasting research, not evidence of a reliable trading service. No general accuracy, financial return or production-readiness claim follows from that historical result.