Accuracy Metrics
Every evening after the USD/JPY market closes, the model publishes a directional forecast for the next trading day. The next day's actual close is compared to the forecast and the result is recorded in real time. All predictions are strictly out-of-sample: the model has never seen this data during training.
Overall Accuracy
7d
20.0%
95% CI: 3.6% – 62.4%
1 / 5 correct
30d
47.6%
95% CI: 28.3% – 67.6%
10 / 21 correct
90d
47.6%
95% CI: 28.3% – 67.6%
10 / 21 correct
All time
47.6%
95% CI: 28.3% – 67.6%
10 / 21 correct
How accuracy is calculated — All predictions with a known actual outcome are counted — Bull, Bear, and Uncertain. A prediction is correct when the model's signal matches the actual outcome for that day.
What the confidence range means — The range is a 95% Wilson confidence interval. It shows the plausible range for the model's true long-run accuracy given the available data. A wide range means too few predictions to draw firm conclusions — it narrows naturally as more forecasts accumulate.
Confusion Matrix
| Actual ↓ · Pred →↓ / → | Bull | Bear | Uncertain |
|---|---|---|---|
| Bull | 8 | 0 | 0 |
| Bear | 0 | 0 | 6 |
| Uncertain | 5 | 0 | 2 |
Precision & Recall
| Signal | Precision | Recall |
|---|---|---|
| Bull | 61.5% | 100.0% |
| Bear | — | 0.0% |
| Uncertain | 25.0% | 28.6% |
Rows show what actually happened. Columns show what the model predicted. Green numbers on the diagonal = correct calls. Off-diagonal = mistakes.
Precision — Of all days the model called Bull, what share were actually bullish? High precision means few false alarms.
Recall — Of all days that were actually bullish, what share did the model catch? High recall means few missed moves.
Last updated: 2026-08-18T03:09:33Z