Evaluating Stock-Market Forecasts in India

Connected evidence: Compare this analysis with SEBI Social Stock Exchange framework for market intermediaries, SEBI Stock Brokers Regulations for market intermediaries, Stock Split definition and Designated Stock Exchange — definition and examples.

Accuracy is not one metric

A forecast can estimate direction, return, volatility, a range or a probability. Each target requires a different evaluation measure. Directional hit rate should be compared with the majority-class baseline. Return estimates need absolute or squared error. Probability forecasts need calibration and a proper scoring rule. Publishing a single accuracy percentage without the target, horizon and sample size prevents meaningful review.

Use the information available at the time

Point-in-time testing reconstructs what was knowable when the forecast was issued. Index constituents, company results and macroeconomic releases can be revised. A backtest that uses the latest constituent list or a later-restated figure leaks future information. The forecast record should preserve input snapshots, model version, issue time, evaluation time and any abstention.

A minimum public scorecard

• Target, universe, forecast horizon and benchmark.
• Coverage, missing forecasts and abstention rate.
• Directional accuracy plus magnitude error where applicable.
• Calibration by probability bucket and market regime.
• Results before and after estimated costs, with uncertainty intervals.

Compare like with like

Technical, fundamental, macroeconomic and machine-learning models often answer different questions. The fair comparison uses the same universe, information cut-off, horizon and evaluation rule. A complex model earns its place only if it adds stable value over transparent baselines after costs and across out-of-sample periods. A result from one regime should not be generalized without evidence.

Use boundary

Forecast evaluation describes historical model behaviour, not a promise about future returns. ShareKeyX uses scorecards to expose uncertainty and failure modes. Readers should combine this evidence with independent research and their own risk constraints.