Commercial Crude Inventory Forecast

Walk-forward holdout

Weekly change in US commercial crude stocks, one week ahead · 104-week out-of-sample evaluation

Mean absolute error

2,562kb

104-wk holdout-1,424 kb vs baseline

Root mean squared error

3,230kb

RMSE/MAE 1.26×penalises big misses

Explained variance

Useful fit

0.270

R² on the weekly change

Directional hit rate

72.1%build/draw

vs random (50%)+22.1 pts

Actual vs model forecast

Weekly stock change in thousand barrels, rolling out-of-sample with a 90% band

Actual EIA
LightGBM
90% CI

Specification

LightGBM · 400 trees · depth 6 · 45 features

AIC

n/a for this model

Std error

±3,230 kb

Holdout

104 weeks

Gemini AI Assistant

Preview · not connected

Imports carry the most weight
2m ago
Across the holdout, the 4-week import average has the strongest positive effect on the weekly change.
Residuals look well behaved
18m ago
Ljung-Box finds no autocorrelation left in the errors, and Jarque-Bera does not reject normality.

Top drivers

Standardised effect on the weekly inventory change

[−1, +1]

Refinery utilisation (lag 1)p = 0.001−0.36
Imports, 4-week averagep = 0.001+0.29
Exports (lag 1)p = 0.003−0.24
Seasonal week of yearp = 0.005−0.17
Production (lag 1)p = 0.020+0.11
WTI–Brent spreadp = 0.090−0.07

Left: pulls stocks downRight: builds stocks

Diagnostics & residuals

Stationary (ADF pass)

The target is the weekly change rather than the stock level, so the series is already differenced. Residuals are checked for leftover autocorrelation and for normality before the uncertainty band is trusted.

Ljung-Box Q (p-value): 0.538 (no autocorrelation)
Jarque-Bera (normality): 1.52 (normal)
Validation: Walk-forward, 104 weeks

Model comparison

All candidates on the same 104-week walk-forward holdout. Lower MAE and RMSE are better.

Model	MAE (kb)	RMSE (kb)	R²	Direction	vs seasonal naive
Naive	3,978	5,051	-0.786	54.8%	+0.2%
Seasonal Naive	3,985	4,923	-0.696	49.0%	baseline
SARIMAX	3,121	3,778	0.001	65.4%	+21.7%
LightGBMSelected	2,562	3,230	0.270	72.1%	+35.7%
LSTM	3,086	3,862	-0.044	67.3%	+22.6%