Accuracy
How well does it work?
The numbers behind the forecasts. The real-data backtest is the one to trust — the model is tested against real prices it never saw during training.
Regression — synthetic train / test split
| Model |
MAE |
RMSE |
MAPE |
R² |
| Naive (expected_price) |
$88.01 |
$190.86 |
21.18% |
0.8170 |
| Linear Regression |
$93.61 |
$190.05 |
24.98% |
0.8185 |
| Decision Tree |
$89.08 |
$193.52 |
21.86% |
0.8119 |
| Random Forest |
$88.85 |
$193.85 |
21.41% |
0.8112 |
The naive baseline (expected_price alone) sits within a dollar of the best ML model. Most of the R² comes from the linear interpolation between 2023 and 2025 anchors, not from features the model learned.
Per-Category MAE
| Model | CPU | GPU | RAM | Storage |
|---|
| Decision Tree |
$9.29 | $127.77 | $107.76 | $71.60 |
| Linear Regression |
$10.29 | $158.52 | $106.27 | $66.85 |
| Naive (expected_price) |
$7.52 | $127.54 | $107.73 | $69.44 |
| Random Forest |
$9.03 | $128.57 | $107.90 | $70.52 |
Real-Data Out-of-Time Backtest
Train on synthetic plus the early real-blended months. Test on the held-out last three months of real prices — values the model never saw. This is the credible accuracy claim.
GPU
| Model |
MAE |
RMSE |
MAPE |
R² |
n test |
| Linear Regression |
$327.86 |
$486.70 |
41.52% |
0.7301 |
559 |
| Decision Tree |
$218.31 |
$387.78 |
23.00% |
0.8286 |
559 |
| Random Forest |
$211.94 |
$361.51 |
23.17% |
0.8511 |
559 |
Storage
| Model |
MAE |
RMSE |
MAPE |
R² |
n test |
| Linear Regression |
$125.96 |
$209.82 |
45.41% |
0.5056 |
1,689 |
| Decision Tree |
$88.59 |
$140.47 |
30.44% |
0.7784 |
1,689 |
| Random Forest |
$61.82 |
$106.61 |
21.90% |
0.8723 |
1,689 |
RAM
| Model |
MAE |
RMSE |
MAPE |
R² |
n test |
| Linear Regression |
$197.16 |
$278.40 |
55.73% |
0.1242 |
1,055 |
| Decision Tree |
$69.71 |
$98.71 |
24.17% |
0.8899 |
1,055 |
| Random Forest |
$55.73 |
$97.14 |
15.69% |
0.8934 |
1,055 |
Permutation Feature Importance
Drop in negative-MAE when each feature is shuffled. expected_price does almost all the work — shuffling any other feature barely moves the error.
| Feature | Importance (mean) | Std |
|---|
| expected_price |
242.3123 |
0.3293 |
| cat_RAM |
0.3813 |
0.0179 |
| sin_month |
0.0711 |
0.0071 |
| cos_month |
0.0580 |
0.0031 |
| trend_per_step |
0.0000 |
0.0000 |
| event_active |
0.0000 |
0.0000 |
| cat_CPU |
-0.0521 |
0.0080 |
| spec_value |
-0.3145 |
0.0088 |
Hardest Test Predictions
The components the model got most wrong — usually exotic high-end SKUs with thin training data.
| Category | Model |
Actual | Predicted | Abs error |
| Storage |
intel dc s3700 |
$1502.42 |
$434.15 |
$1086.38 |
| GPU |
pny founders edition |
$1227.80 |
$399.44 |
$897.16 |
| RAM |
adata xpg caster rgb 32 gb |
$1198.30 |
$510.75 |
$716.44 |
| GPU |
evga xc ultra gaming |
$1140.18 |
$579.77 |
$622.85 |
| GPU |
msi armor oc |
$1222.36 |
$687.00 |
$610.14 |
| GPU |
asus rog strix lc gaming oc |
$1337.86 |
$1891.92 |
$554.07 |
| GPU |
asus strix gaming |
$1255.38 |
$810.29 |
$553.48 |
| GPU |
gigabyte aorus master |
$1134.81 |
$1683.08 |
$548.26 |
| GPU |
msi founders edition |
$1164.54 |
$737.65 |
$534.48 |
| GPU |
zotac amp extreme holo |
$1490.18 |
$2023.44 |
$533.74 |
Cross-Sectional Spec → Price (real prices)
Train on 2023 (component, price) pairs; test on held-out 2025 pairs. Both train and test labels are real prices — no synthetic data involved.
| Category | Model |
MAE | RMSE | MAPE | R² |
| GPU |
Linear Regression |
$247.65 |
$386.09 |
34.04% |
0.6114 |
| GPU |
Decision Tree |
$213.97 |
$328.16 |
24.69% |
0.7192 |
| GPU |
Random Forest |
$210.19 |
$322.90 |
24.33% |
0.7282 |
| CPU |
Linear Regression |
$116.31 |
$173.44 |
74.83% |
0.3292 |
| CPU |
Decision Tree |
$93.48 |
$158.08 |
56.63% |
0.4428 |
| CPU |
Random Forest |
$91.22 |
$148.58 |
55.23% |
0.5077 |
| RAM |
Linear Regression |
$44.76 |
$74.04 |
47.59% |
0.6608 |
| RAM |
Decision Tree |
$41.13 |
$74.92 |
34.25% |
0.6527 |
| RAM |
Random Forest |
$39.99 |
$70.86 |
34.76% |
0.6893 |
| Storage |
Linear Regression |
$86.61 |
$134.61 |
85.18% |
0.2466 |
| Storage |
Decision Tree |
$73.07 |
$117.40 |
60.23% |
0.4269 |
| Storage |
Random Forest |
$73.54 |
$117.54 |
59.94% |
0.4255 |
Spec → Tier Classifier
| Category | Model |
Accuracy | Macro F1 | Macro P | Macro R |
| GPU |
Logistic Regression |
0.763 |
0.763 |
0.764 |
0.762 |
| GPU |
Decision Tree |
0.811 |
0.808 |
0.815 |
0.810 |
| GPU |
Random Forest |
0.833 |
0.831 |
0.832 |
0.833 |
| CPU |
Logistic Regression |
0.636 |
0.612 |
0.644 |
0.636 |
| CPU |
Decision Tree |
0.688 |
0.660 |
0.682 |
0.688 |
| CPU |
Random Forest |
0.719 |
0.702 |
0.722 |
0.719 |
| RAM |
Logistic Regression |
0.765 |
0.766 |
0.770 |
0.765 |
| RAM |
Decision Tree |
0.773 |
0.768 |
0.781 |
0.773 |
| RAM |
Random Forest |
0.781 |
0.778 |
0.787 |
0.781 |
| Storage |
Logistic Regression |
0.624 |
0.602 |
0.649 |
0.624 |
| Storage |
Decision Tree |
0.651 |
0.635 |
0.678 |
0.650 |
| Storage |
Random Forest |
0.669 |
0.673 |
0.689 |
0.669 |