TREND / NOW
Trend assessment
Model 1 describes current market conditions. Trend scores and factor consistency are not the probability of a rise over the next 24 hours.
Review existing trend assessments ↗MODEL 03 / INDEPENDENT FORECAST RESEARCH
Gold
Statistical forecast researchLet time test the forecast.
Independently forecast the next 24-hour direction without treating the current trend score as a forecast probability. Train regularized logistic regression on historical market data; store and validate its results separately from Models 1 and 2.
View full validation ↓No advantage over simple baselines has been established. Do not use holdout results to reselect parameters or present a partial hit rate with low coverage as meeting the overall target. Existing trend pages and daily updates remain unchanged.
FORECAST / NEXT 24H
Independent direction forecast
Model 3 learns the relationship between historical features and the subsequent 24-hour direction. Raw model outputs are not yet calibrated and are not promises of success probabilities.
Today's forecast: not enabledFORWARD / FROZEN EVIDENCE
Post-publication validation
First freeze evidence of the initial publication, then wait 24 hours for validation. Do not backfill historical simulations as live records.
Published 0 times · Verified 0 times · Hit rate —01 / CHRONOLOGICAL HOLDOUT
Hold out first, then verify
Retain the last 40 scheduled days; days with missing prices remain in the scheduled total.
Gold: hit · Gray: miss · Time runs left to right. The primary test threshold is fixed at 0.50, with full coverage. Retraining for the latter 20 days may use earlier labels that have matured by then; parameters are not reselected.
Key limitation: this is a chronological historical simulation, not a genuinely blind test of data never seen by the researchers. Some recent history had already been examined while investigating PR #21; a fresh forward-validation period is still needed. Wilson intervals do not adjust for temporal dependence and are only a reference for sampling uncertainty.
02 / SAME WINDOW BASELINES
Compare against simple methods first
All use the same 40-day window; no improvement claims across different windows.
| Same holdout window | Hits / Samples | Directional hit rate | Brier score ↓ |
|---|---|---|---|
| Model 3 · Logistic regression | 19 / 40 | 47.5% | 0.2602 |
| Always bullish | 18 / 40 | 45.0% | 0.5500 |
| Always bearish | 22 / 40 | 55.0% | 0.4500 |
| Previous-day momentum | 14 / 40 | 35.0% | 0.6500 |
| Reversed momentum | 26 / 40 | 65.0% | 0.3500 |
| Training-period up/down prevalence | 22 / 40 | 55.0% | 0.2494 |
“Training-period up/down prevalence” uses only training labels that had matured at the time. Momentum and reversed momentum are prespecified baselines only; they cannot be selected for deployment simply after viewing holdout results. The 0/1 outputs of deterministic directional baselines are not calibrated probabilities.
03 / SIGNAL COVERAGE
Hit rates must be read alongside coverage
0.50 is the primary test; all other thresholds are listed for diagnostic purposes only.
| Raw model output threshold | Hits / Samples | Directional hit rate | Signals issued / Scheduled days | Coverage |
|---|---|---|---|---|
| 0.50 · Primary test | 19 / 40 | 47.5% | 40 / 40 | 100.0% |
| 0.60 | 7 / 12 | 58.3% | 12 / 40 | 30.0% |
| 0.70 | 1 / 2 | 50.0% | 2 / 40 | 5.0% |
| 0.80 | 0 / 0 | — | 0 / 40 | 0.0% |
| 0.90 | 0 / 0 | — | 0 / 40 | 0.0% |
Threshold t: raw p ≥ t is bullish, 1−p ≥ t is bearish; otherwise no call. With no signals issued, the hit rate is “—”, not 100%. A higher threshold does not imply validated high confidence.
04 / METHOD & SELECTION
Selection rules precede holdout results
Select parameters only by development-set Log loss, not by choosing the best-looking hit-rate table.
Training and information timing
Feature cutoff: closed candles from the previous UTC day. The forecast cutoff is 1 minute before the starting-price minute; training labels must end strictly before that cutoff. Training means and standard deviations are calculated only within the training block; transformed values are capped at ±5.
80–240 training samples; up to 3 folds at the end of the development period, each covering 20 scheduled days. L2 is 0.1 / 1 / 10; 800 deterministic gradient-descent steps, with no penalty on the intercept.
Features and market boundaries
Core: 1/3/7-day momentum, EMA spread, RSI and 20-day range position. Extended: volatility regime, closing position, volume and range amplitude. Final selection: extended / L2 10.
Gold uses Binance XAUUSDT perpetuals, not London spot XAU/USD. Input days are Monday–Friday UTC (excluding dates in the existing holiday calendar), with forecasts starting at 00:15 the following day; the 24-hour wall-clock window includes weekends.
| Feature set / L2 | Rolling validation samples | Validation Log loss ↓ | Validation hit rate | Selection status |
|---|---|---|---|---|
| core / 0.1 | 40 | 0.7861 | 42.5% | Not selected |
| core / 1 | 40 | 0.7756 | 42.5% | Not selected |
| core / 10 | 40 | 0.7451 | 35.0% | Not selected |
| extended / 0.1 | 40 | 0.7721 | 47.5% | Not selected |
| extended / 1 | 40 | 0.7579 | 47.5% | Not selected |
| extended / 10 | 40 | 0.7313 | 45.0% | Lowest validation loss · Locked |
Model output versus observed share of rises (not proof of successful calibration)
| Model output range | Sample count | Mean raw p(up) | Observed share of rises |
|---|---|---|---|
| 0.0 — 0.2 | 0 | — | — |
| 0.2 — 0.4 | 9 | 0.3549 | 33.3% |
| 0.4 — 0.6 | 28 | 0.4997 | 50.0% |
| 0.6 — 0.8 | 3 | 0.6713 | 33.3% |
| 0.8 — 1.0 | 0 | — | — |
05 / ROW-BY-ROW AUDIT
Every record can be reviewed
Prices are opening prices of 1-minute candles from the same source; transaction costs are excluded and profits are not simulated.
Expand 40 historical forecasts and training-time evidence
| Input day / UTC start | Raw p(up) | Simulated assessment | Starting price → 24h later | Verification | Latest training-label endpoint UTC |
|---|---|---|---|---|---|
| 2026-08-202026-08-21 00:15 | 0.5845 | Bullish | 4,522.08 → 4,616.3 | Hit | 2026-08-20 00:15 |
| 2026-08-212026-08-22 00:15 | 0.5590 | Bullish | 4,616.3 → 4,609.29 | Miss | 2026-08-20 00:15 |
| 2026-08-242026-08-25 00:15 | 0.3591 | Bearish | 4,693.93 → 4,654.03 | Hit | 2026-08-20 00:15 |
| 2026-08-252026-08-26 00:15 | 0.4997 | Bearish | 4,654.03 → 4,620.45 | Hit | 2026-08-20 00:15 |
| 2026-08-262026-08-27 00:15 | 0.3675 | Bearish | 4,620.45 → 4,603.57 | Hit | 2026-08-20 00:15 |
| 2026-08-272026-08-28 00:15 | 0.3933 | Bearish | 4,603.57 → 4,467.99 | Hit | 2026-08-20 00:15 |
| 2026-08-282026-08-29 00:15 | 0.4120 | Bearish | 4,467.99 → 4,473.82 | Miss | 2026-08-20 00:15 |
| 2026-08-312026-09-01 00:15 | 0.3909 | Bearish | 4,459.8 → 4,335.59 | Hit | 2026-08-20 00:15 |
| 2026-09-012026-09-02 00:15 | 0.4248 | Bearish | 4,335.59 → 4,394.99 | Miss | 2026-08-20 00:15 |
| 2026-09-022026-09-03 00:15 | 0.4698 | Bearish | 4,394.99 → 4,483.82 | Miss | 2026-08-20 00:15 |
| 2026-09-032026-09-04 00:15 | 0.5177 | Bullish | 4,483.82 → 4,435.81 | Miss | 2026-08-20 00:15 |
| 2026-09-042026-09-05 00:15 | 0.5369 | Bullish | 4,435.81 → 4,437.07 | Hit | 2026-08-20 00:15 |
| 2026-09-072026-09-08 00:15 | 0.4464 | Bearish | 4,420.78 → 4,354.89 | Hit | 2026-08-20 00:15 |
| 2026-09-082026-09-09 00:15 | 0.4349 | Bearish | 4,354.89 → 4,398.63 | Miss | 2026-08-20 00:15 |
| 2026-09-092026-09-10 00:15 | 0.4680 | Bearish | 4,398.63 → 4,324.14 | Hit | 2026-08-20 00:15 |
| 2026-09-102026-09-11 00:15 | 0.4775 | Bearish | 4,324.14 → 4,357.5 | Miss | 2026-08-20 00:15 |
| 2026-09-112026-09-12 00:15 | 0.5208 | Bullish | 4,357.5 → 4,357.29 | Miss | 2026-08-20 00:15 |
| 2026-09-142026-09-15 00:15 | 0.4544 | Bearish | 4,291.23 → 4,288.54 | Hit | 2026-08-20 00:15 |
| 2026-09-152026-09-16 00:15 | 0.3996 | Bearish | 4,288.54 → 4,291.26 | Miss | 2026-08-20 00:15 |
| 2026-09-162026-09-17 00:15 | 0.2702 | Bearish | 4,291.26 → 4,361.2 | Miss | 2026-08-20 00:15 |
| 2026-09-172026-09-18 00:15 | 0.5096 | Bullish | 4,361.2 → 4,384.05 | Hit | 2026-09-17 00:15 |
| 2026-09-182026-09-19 00:15 | 0.5039 | Bullish | 4,384.05 → 4,380.59 | Miss | 2026-09-17 00:15 |
| 2026-09-212026-09-22 00:15 | 0.4917 | Bearish | 4,374.8 → 4,369.04 | Hit | 2026-09-17 00:15 |
| 2026-09-222026-09-23 00:15 | 0.2791 | Bearish | 4,369.04 → 4,292.78 | Hit | 2026-09-17 00:15 |
| 2026-09-232026-09-24 00:15 | 0.5288 | Bullish | 4,292.78 → 4,270.25 | Miss | 2026-09-17 00:15 |
| 2026-09-242026-09-25 00:15 | 0.5104 | Bullish | 4,270.25 → 4,292.18 | Hit | 2026-09-17 00:15 |
| 2026-09-252026-09-26 00:15 | 0.5221 | Bullish | 4,292.18 → 4,286.87 | Miss | 2026-09-17 00:15 |
| 2026-09-282026-09-29 00:15 | 0.5773 | Bullish | 4,127.75 → 4,187.77 | Hit | 2026-09-17 00:15 |
| 2026-09-292026-09-30 00:15 | 0.5908 | Bullish | 4,187.77 → 4,157.46 | Miss | 2026-09-17 00:15 |
| 2026-09-302026-10-01 00:15 | 0.6913 | Bullish | 4,157.46 → 4,174.37 | Hit | 2026-09-17 00:15 |
| 2026-10-012026-10-02 00:15 | 0.6324 | Bullish | 4,174.37 → 4,149.61 | Miss | 2026-09-17 00:15 |
| 2026-10-022026-10-03 00:15 | 0.6901 | Bullish | 4,149.61 → 4,145.21 | Miss | 2026-09-17 00:15 |
| 2026-10-052026-10-06 00:15 | 0.4570 | Bearish | 4,139.67 → 4,172.26 | Miss | 2026-09-17 00:15 |
| 2026-10-062026-10-07 00:15 | 0.3767 | Bearish | 4,172.26 → 4,114.77 | Hit | 2026-09-17 00:15 |
| 2026-10-072026-10-08 00:15 | 0.4841 | Bearish | 4,114.77 → 4,151.59 | Miss | 2026-09-17 00:15 |
| 2026-10-082026-10-09 00:15 | 0.5143 | Bullish | 4,151.59 → 4,195.78 | Hit | 2026-09-17 00:15 |
| 2026-10-092026-10-10 00:15 | 0.5426 | Bullish | 4,195.78 → 4,195.69 | Miss | 2026-09-17 00:15 |
| 2026-10-122026-10-13 00:15 | 0.4617 | Bearish | 4,167.5 → 4,169.09 | Miss | 2026-09-17 00:15 |
| 2026-10-132026-10-14 00:15 | 0.4906 | Bearish | 4,169.09 → 4,136.08 | Hit | 2026-09-17 00:15 |
| 2026-10-142026-10-15 00:15 | 0.3578 | Bearish | 4,136.08 → 4,161.77 | Miss | 2026-09-17 00:15 |
06 / HORIZON COMPARISON
Which horizon merits further validation?
An additional exploratory experiment; it does not replace the original experiment above with better-looking results.
| Forecast horizon | Validation-set hit rate | Holdout hits / Samples | Holdout hit rate | Coverage | 7-day block interval (95%) |
|---|---|---|---|---|---|
| 1 hour | 52.4% | 238 / 480 | 49.6% | 100.0% | 43.3% — 55.4% |
| 4 hours | 48.3% | 55 / 120 | 45.8% | 100.0% | 35.8% — 55.0% |
| 8 hours | 52.3% | 26 / 60 | 43.3% | 100.0% | 33.3% — 53.3% |
| 12 hours | 41.4% | 22 / 40 | 55.0% | 100.0% | 37.5% — 70.0% |
| 24 hours · Preselected on the validation set | 57.1% | 12 / 20 | 60.0% | 100.0% | 45.0% — 75.0% |
Same holdout period: 2026-09-18 to 2026-10-16 UTC (right endpoint excluded), a 90-day training window, retraining every 7 days. Anchored at midnight UTC, each forecast predicts the next H hours and is issued every H hours; target labels do not overlap within a horizon, and features use closed H-hour candles. This clock, training window and 28-day test period differ from the initial 40-day experiment above and cannot be combined for comparison.
Horizons are preselected only on the validation set; do not choose a new winner after viewing holdout results. Validation selected 24 hours, with holdout performance of 60.0% (12/20), exactly matching the always-bearish baseline of 60.0%. The lower bounds of the 7-day block intervals for all horizons are below 55%, and exploration of multiple horizons has no multiple-comparison correction; no horizon has yet been shown to meet the target consistently.
Expand same-window baselines for each horizon
| Cycle | Model | Always bullish | Always bearish | Previous-period momentum | Reversed momentum |
|---|---|---|---|---|---|
| 1 hour | 49.6% | 48.8% | 51.0% | 49.2% | 50.6% |
| 4 hours | 45.8% | 49.2% | 50.8% | 52.5% | 47.5% |
| 8 hours | 43.3% | 58.3% | 41.7% | 53.3% | 46.7% |
| 12 hours | 55.0% | 42.5% | 57.5% | 55.0% | 45.0% |
| 24 hours | 60.0% | 40.0% | 60.0% | 45.0% | 55.0% |
07 / NEXT STEP
Review first, then gather new evidence
This page does not enable automatic publication or operation.
- Grok reviews the code, splitting rules, source markets and all backtest records.
- Register new experiments before validating future versions; the current holdout has already been used and cannot be reused as a fresh blind test.
- After user approval, Grok alone executes daily updates; the three sets of data, publication evidence and hit rates remain independent, and old records are not rewritten.
55% is the first-stage research target, not a promise of returns; 90% is not this round's acceptance criterion. Live forward validation has not started, so the target cannot be claimed as achieved.
Research snapshot as of 2026-10-16 20:24:44 UTC · Model model-3-logistic-v1 · schemaVersion 1
Data snapshot SHA-256 c1b157596464446515e1f43afa2cfd76837d3a73df9d3852b90129448e680a8a
Protocol SHA-256 cfb6f3da2dc9e7db066451eb1161113d5ce429ccb1c5be3e23cef61a3787633c
Complete public market snapshots, fitted parameters, row-by-row records and reproduction instructions are committed with the PR. This page is a research snapshot fixed at build time; it does not read, backfill or modify the three live daily data feeds. For research only; not investment advice.
View Bitcoin Model 3 research ↗