AR error correction: one-step MAE -33%, direction accuracy +6.6pp, Day 16 bias eliminated
One-step MAE dropped 33% in a single experiment. A persistent $2.99 pricing error on Day 16 : visible in every backtest : was reduced to $0.05. The fix was autoregressive error correction.
Context
After the tail risk recalibration in v13, the oil model had a different problem: a persistent bias on Day 16 of each forecast cycle. Day 16 consistently showed a ~$3 underestimation. The error was not random : it appeared in backtests across different market conditions. This kind of systematic bias suggests the model is missing a structural feature rather than just needing better parameters.
The v14 experiment investigated whether autoregressive error correction (learning from prior forecast errors) could eliminate this Day 16 artifact while improving overall accuracy.
What Changed
Two mechanisms were added. First, autoregressive error correction with arLambda=0.4: the model now tracks its own prior-period errors and applies a weighted correction to the current forecast. A parameter of 0.4 means 40% of the prior error is corrected in the next step. Second, spike reversion mechanics were added to handle the tail cases where a large single-period price spike doesn’t persist : the model now incorporates mean-reversion pressure after detected spikes.
As a side effect, the model was also simplified: active parameters were reduced from 16 to 8 by removing redundant tuning variables that the AR correction now rendered unnecessary. Fewer parameters means lower overfitting risk on future calibrations.
Impact
MAE dropped from $1.63 to $1.09 : a 33% reduction in absolute forecast error.
Before: One-step MAE $1.63, direction accuracy 86.7%, Day 16 error -$2.99, 16 active parameters. After: One-step MAE $1.09, direction accuracy 93.3% (+6.6pp), Day 16 error +$0.05 (essentially eliminated), 8 active parameters.
The Day 16 bias dropping from -$2.99 to +$0.05 is the clearest result: a systematic error that appeared in every backtest was eliminated by targeting the structural cause rather than patching the symptom.