SAHARAN BLEND · 2000–2022 · 7 MODELS BENCHMARKED

Crude Oil Price Forecast — Model Performance

Seven regression and deep learning models were trained to predict the Saharan Blend price, evaluated on held-out test data across two granularities. Switch between daily and monthly data — hover the chart for exact values.

Observations: ≈ 8,385 daily observations

Split: 64% train · 16% val · 20% test

Target: Saharan Blend price ($/bbl)

Best model

RF-XGB — RMSE 0.24 · R² 1.00

Comparison of Regression Models — Daily data

R² (bars) RMSE (line)

Full results table

Model	RMSE ($/bbl)	R²
SVR	21.81	0.32
RF	18.52	0.51
XGB	19.82	0.44
CNN	5.29	0.96
LSTM	10.41	0.72
CNN-LSTM	0.56	1.00
RF-XGB ★	0.24	1.00

On high-frequency daily data, sequence-aware and hybrid models dominate. Both hybrids (CNN-LSTM and RF-XGB) reach near-perfect fit, but RF-XGB edges ahead with the lowest RMSE ($0.24/bbl). Classical SVR and untuned tree models struggle with the daily noise level, R² capped around 0.3–0.5.

Observations: 276 monthly observations

Split: 64% train · 16% val · 20% test

Target: Saharan Blend price ($/bbl)

Best model

CNN-LSTM — RMSE 0.21 · R² 1.00

Comparison of Regression Models — Monthly data

R² (bars) RMSE (line)

Full results table

Model	RMSE ($/bbl)	R²
SVR	21.69	0.33
RF	18.50	0.51
XGBoost	19.83	0.44
LSTM	2.01	0.99
CNN	1.58	1.00
CNN-LSTM ★	0.21	1.00
RF-XGB	19.44	0.46

On the monthly dataset, the pattern flips: deep learning models (LSTM, CNN, CNN-LSTM) achieve near-perfect fits with R² ≥ 0.99, and CNN-LSTM takes the top spot with an RMSE of just 0.21. Classical SVR and the RF-XGBoost hybrid trail well behind here, staying under R² = 0.5.