Historical research · April 2025

Forecast accuracy didn’t become trading value

The saved directional strategy returned -22.22%, against 15.66% for buy-and-hold. That is a 37.88 percentage-point shortfall.

Every number below comes from the result files committed to the coal forecasting repository. No data is downloaded, no model is refit and no sample observations are created.

The saved price and forecast series

Saved priceSaved forecast

Source: directional strategy detail CSV, Price and Forecast columns.

Read the plotted values

Month	Price	Forecast
2023-01-31	126.1117	113.9555
2023-02-28	95.1977	93.0315
2023-03-31	104.3325	103.5090
2023-04-30	127.8465	125.0873
2023-05-31	136.6089	134.4909
2023-06-30	128.7407	128.5506
2023-07-31	125.0738	124.6027
2023-08-31	122.9303	123.6102
2023-09-30	105.9073	105.3750
2023-10-31	117.9257	115.7191
2023-11-30	122.4521	123.9321
2023-12-31	145.8568	141.4906

Model fit and test errors

Metric	Training	Test
R2	-0.9388	0.3302
RMSE	18.9249	11.0654
MAE	10.0590	9.4886
MAPE	9.50%	7.66%

From the saved model metrics. The historical report defines training as 2020–2022 and testing as 2023. Negative training R² deserves attention alongside the lower test error.

Trading simulation

Strategy	Total return	Buy-and-hold
Directional	-22.22%	15.66%
Threshold	0.00%	15.66%
Combined	0.00%	15.66%

The saved threshold and combined evaluations took no positions. These are simulated results from the repository, with no claim of realized investment returns.

Inspect the source files

File hashes and source paths

A forecast earns a trading claim only after it beats a relevant benchmark after costs. The saved directional strategy failed that test.