← Back to gold AI investment research (original assessment page)
Gold · Algorithm B: 24-hour direction forecast
A deterministic statistical model with no AI calls. Independent of the trend assessment on the original AI investment research page; the “historical backtest (not live)” and “live post-publication validation” below are measured separately and cannot substitute for each other.
Today's 24-hour forecast
Loading forecast
—
—
Historical backtest (not live)Out of sample · Excludes publication delay
Loading backtest
—
—
- Bullish hits
- —
- Bearish hits
- —
- Baselines: always bullish / Previous day's direction
- —
- Same-period results on the verification instrument
- —
Directional hit rate 24 hours after publication
Loading statisticsObserve prices from the first confirmed public publication to validate Algorithm B's direction frozen at publication (days with no forecast are excluded).
—Hits ÷ (Hits + Misses)
Sample pending loadingHit count awaiting retrieval
- Awaiting a full 24 hours
- —
- Missing publication evidence
- —
- Fully neutral
- —
- Market data requires verification
- —
Statistics cover all archives included in the validation materials; date filters do not change this statistic.
Statistics update time awaiting retrieval
Statistical definitions and validation boundaries
- Forecast direction:Only use trend scores frozen at first publication; above 0 is bullish and below 0 bearish, including “neutral with a bullish/bearish bias”; purely neutral scores equal to 0 are excluded. Recalculated assessments cannot replace the originally published direction.
- Published:The conservative confirmed public publication time is when the first upload is read back from the public internet, which may be later than the actual upload. Archived trading dates, data assembly logs, AI completion logs and file modification times cannot substitute for it. Historical archives without reliable confirmation records are neither backfilled nor included.
- Observation window:The start is the opening price of the first 1-minute candle no earlier than the confirmed publication time, rounded forward with a delay under 60 seconds; the endpoint is the opening price of the corresponding minute 24 hours later. Daily changes or the following day's close are not substituted.
- Validation market data:Binance XAUUSDT gold perpetual futures (USDT/ounce), not spot gold; a continuous 24-hour period including weekends, without rolling forward to the next trading day. Uses only prices from the same market and basis above; unavailable sources or missing prices are labeled “Market data requires verification,” without substitution from live-quote fallback sources.
- Hit assessment:A bullish view with an ending price above the starting price, or a bearish view with an ending price below it, counts as a hit; the opposite or unchanged prices count as a miss. Valid samples include only “hits + misses”; cases under 24 hours, missing publication evidence, purely neutral views and missing market data are excluded.
- Statistical meaning:This validates the trend score's direction; it does not assess AI text quality or equal a trading win rate or investment return. Fees, slippage and holding paths are excluded. Small samples can cause volatile ratios, and past hits do not guarantee future performance.
View original validation records ↗Methodology version publication-24h-v1
Methodology
- Decision time:00:00 UTC on the day following the UTC trading day. Uses only 1-hour candles closed before that time, with daily publication at approximately 00:1x UTC and no look-ahead data.
- Features:13 features: multi-horizon log returns from 4h–30d, three EMA ratios, daily RSI14, position within the 20-day range, and the short-/long-term volatility ratio.
- Model:L2-regularized logistic regression, refitted daily using all historical samples whose target windows have ended by decision time; training data are PAXGUSDT spot 1h (from 2020, using tokenized gold as a long-history proxy).
- Abstention:No forecast is issued when the probability of a rise differs from 50% by no more than ±1%; these cases are excluded from the hit rate. The threshold is selected using only data before the out-of-sample period begins.
- Backtest:Rolling refitting every 30 days during the out-of-sample period, calculating 24-hour direction from 1-hour opening prices; flat prices count as misses. Live validation instead uses Binance XAUUSDT perpetual prices (including weekends; not spot XAU/USD) at minute boundaries aligned after confirmed publication.
- Limitations:The out-of-sample hit rate's 95% interval includes 50%; superiority over chance has not been statistically established. Not investment advice.
FACTS / SOURCES / CONTEXT
Page highlights and supporting materials
What price basis does gold research use?
Historical research primarily uses Binance XAUUSDT perpetual contract data, with OKX XAU-USDT-SWAP as fallback; specific sources are recorded in each period's data. Check the selected UTC trading day. Contract data are not equivalent to physical gold trades or gold tailings project selling prices.