Analysis — Direction
Recomputed from btc.db on every request · 2026-11-18 15:12:02 UTC
The phase filter applies to all sections. Only evaluated UP/DOWN rows are included. Strength = max(p_up, p_down). Accuracy uses stored correct (missing values are recomputed; ties predict UP). Brier = mean((p_up − outcome)²), where UP = 1 and DOWN = 0; lower is better (50/50 baseline: 0.25). Empty groups show —. Invalid probability rows excluded: 0; correct values recomputed: 0.
Phase comparison
| Phase | Model | N | Correct | Accuracy | Brier score | Avg prediction strength |
|---|---|---|---|---|---|---|
| Phase 2 | GPT | 48 | 23 | 47.9% | 0.2553 | 55.0% |
| Phase 2 | Jev | 48 | 23 | 47.9% | 0.3386 | 76.2% |
Direction accuracy by realized volatility
Retrospective / conditional on realized volatility: subsequent-hour RV is known only afterward. This is exploratory and not tradable prospectively. Phase 4 thresholds were calibrated after the direction experiments. Observations from different models for the same hour are not independent.
RV = 100 × sqrt(sum(log(C_i / C_(i−1))²)), exactly twelve consecutive 5m returns including the preceding close. QUIET: RV < 0.227425610%; NORMAL: 0.227425610% ≤ RV < 0.370664156%; ACTIVE: RV ≥ 0.370664156%.
GPT / OpenAI: ACTIVE 12.5% (N=8), non-ACTIVE 55.0% (N=40); difference -42.5 percentage points. Jev: ACTIVE 50.0% (N=8), non-ACTIVE 47.5% (N=40); difference +2.5 percentage points.
Missing RV excluded: 0; invalid probability rows excluded: 0. ALL includes only RV-assigned rows. Stored actual_return is in percent; missing/invalid returns are excluded only from magnitude metrics (Return N). No interpolation or network fetching occurs in this page.
| Phase | Model | Eligible N | RV assigned | Missing RV |
|---|---|---|---|---|
| phase2 | openai | 48 | 48 | 0 |
| phase2 | jev | 48 | 48 | 0 |
| Phase | Model | Actual regime | N | Correct | Accuracy | Brier | Return N | Mean |return| | Median |return| | Mean RV |
|---|---|---|---|---|---|---|---|---|---|---|
| phase2 | GPT / OpenAI | ALL | 48 | 23 | 47.9% | 0.2553 | 48 | 0.1886% | 0.1550% | 0.2504% |
| phase2 | GPT / OpenAI | QUIET | 30 | 15 | 50.0% | 0.2600 | 30 | 0.1235% | 0.1270% | 0.1635% |
| phase2 | GPT / OpenAI | NORMAL | 10 | 7 | 70.0% | 0.2189 | 10 | 0.2117% | 0.2296% | 0.3022% |
| phase2 | GPT / OpenAI | ACTIVE | 8 | 1 | 12.5% | 0.2831 | 8 | 0.4039% | 0.3847% | 0.5116% |
| phase2 | GPT / OpenAI | NON-ACTIVE | 40 | 22 | 55.0% | 0.2497 | 40 | 0.1455% | 0.1465% | 0.1981% |
| phase2 | Jev | ALL | 48 | 23 | 47.9% | 0.3386 | 48 | 0.1886% | 0.1550% | 0.2504% |
| phase2 | Jev | QUIET | 30 | 12 | 40.0% | 0.3874 | 30 | 0.1235% | 0.1270% | 0.1635% |
| phase2 | Jev | NORMAL | 10 | 7 | 70.0% | 0.2312 | 10 | 0.2117% | 0.2296% | 0.3022% |
| phase2 | Jev | ACTIVE | 8 | 4 | 50.0% | 0.2896 | 8 | 0.4039% | 0.3847% | 0.5116% |
| phase2 | Jev | NON-ACTIVE | 40 | 19 | 47.5% | 0.3484 | 40 | 0.1455% | 0.1465% | 0.1981% |
Calibration analysis
Blue bar: actual accuracy. Black marker: ideal accuracy, equal to the mean prediction strength in that bucket. Buckets include their lower bound and exclude their upper bound, except 100% is included in the last bucket.
GPT
| Strength bucket | N | Accuracy | Ideal (mean strength) | Actual / ideal · 0–100% |
|---|---|---|---|---|
| 50%–55% | 25 | 52.0% | 53.1% | |
| 55%–60% | 21 | 38.1% | 56.7% | |
| 60%–70% | 2 | 100.0% | 62.5% | |
| 70%–80% | 0 | - | - | — |
| 80%–90% | 0 | - | - | — |
| 90%–100% | 0 | - | - | — |
Jev
| Strength bucket | N | Accuracy | Ideal (mean strength) | Actual / ideal · 0–100% |
|---|---|---|---|---|
| 50%–55% | 4 | 25.0% | 52.8% | |
| 55%–60% | 3 | 33.3% | 57.0% | |
| 60%–70% | 6 | 50.0% | 64.8% | |
| 70%–80% | 12 | 66.7% | 74.8% | |
| 80%–90% | 17 | 41.2% | 84.4% | |
| 90%–100% | 6 | 50.0% | 92.5% |
Threshold analysis
Keep predictions whose strength is at least the threshold (not Jev's separate confidence field). Compare accuracy and Brier together with N; higher thresholds leave fewer observations. These are historical subsets, not evidence of future trading profit.
GPT
| Minimum strength | N | Accuracy | Brier score |
|---|---|---|---|
| ≥ 50% | 48 | 47.9% | 0.2553 |
| ≥ 55% | 23 | 43.5% | 0.2612 |
| ≥ 60% | 2 | 100.0% | 0.1409 |
| ≥ 70% | 0 | - | - |
| ≥ 80% | 0 | - | - |
| ≥ 90% | 0 | - | - |
Jev
| Minimum strength | N | Accuracy | Brier score |
|---|---|---|---|
| ≥ 50% | 48 | 47.9% | 0.3386 |
| ≥ 55% | 44 | 50.0% | 0.3459 |
| ≥ 60% | 41 | 51.2% | 0.3518 |
| ≥ 70% | 35 | 51.4% | 0.3650 |
| ≥ 80% | 23 | 43.5% | 0.4420 |
| ≥ 90% | 6 | 50.0% | 0.4357 |
Jev confidence investigation
Probability margin = abs(p_up − p_down). Evaluated Jev rows with valid confidence only. Missing or invalid confidence excluded: 0.
If confidence repeats the probability margin, r approaches 1 and the absolute difference approaches 0. The table groups pairs by margin (lower bound included; 100% included in the last group).
| Probability margin | N | Mean margin | Mean Jev confidence |
|---|---|---|---|
| 0–10% | 4 | 5.5% | 5.8% |
| 10–20% | 4 | 15.5% | 15.0% |
| 20–30% | 1 | 24.0% | 23.0% |
| 30–40% | 5 | 34.8% | 35.2% |
| 40–50% | 4 | 45.0% | 45.2% |
| 50–60% | 7 | 53.7% | 53.6% |
| 60–70% | 9 | 63.6% | 63.6% |
| 70–80% | 8 | 74.5% | 74.5% |
| 80–90% | 6 | 85.0% | 85.0% |
| 90–100% | 0 | - | - |
BTC Predictor