PRISM Forecasting — AI Market Forecasts for Gold, Forex, Indices & Crypto

Benchmark results

We put it against every class of model that matters.

The new wave of AI foundation forecasters — the class of Kronos, Chronos and TimesFM, of which we benchmark the strongest openly runnable, Kronos, head-to-head. The classical quant standards every trading desk still runs — RiskMetrics EWMA, GARCH-family AR volatility, random-walk and rolling-median baselines. And our own previous 54-million-parameter deep system. Identical out-of-sample windows, identical markets, no excuses: PRISM forecasts the entire 48-hour path with the lowest error across the intraday spectrum and wins the overwhelming majority of head-to-head matchups.

PRISM vs classical quant models — full-path error — Full 48-hour path error vs the classical quant standards. Lower is better.

Head-to-head market wins vs Kronos and the previous deep system — Head-to-head wins vs the AI foundation model and our previous deep system.

11 / 11markets won vs Kronos on 4-hour candles

11 / 11markets won vs our previous 54M system on 1-hour candles

−50%error vs the previous production system on 4-hour candles

Model report

Evaluated like a frontier model. Reported like one.

Big-league AI models publish their numbers — so do we. Every figure below is measured on strictly held-out, chronologically future data the model never trained on.

60%peak error reduction vs the no-skill baseline (daily range)

11 / 11markets with a positive edge on 1-hour candles

254exactly-consistent candles per forecast, 4 timeframes

×3independently trained experts vote on every candle

Timeframe	Range MAE	Error reduction*	Close-path MAE	Close-path RMSE
15 minutes	0.079%	−54%	1.04%	1.98%
1 hour	0.147%	−58%	1.05%	2.00%
4 hours	0.307%	−58%	1.10%	2.07%
Daily	0.765%	−60%	1.33%	2.43%

* per-candle range error vs the no-skill baseline, averaged across all 11 markets and the full 48-hour horizon, out-of-sample.

Validation protocol — what professional models require, and what we do:

✓ Chronological walk-forward evaluation ✓ Zero-leakage input audit — every feature provably known at decision time ✓ Out-of-sample-only reporting ✓ Adversarial multi-model benchmarking ✓ Deep ensembling — 3 independent experts ✓ Live re-anchoring every 15 minutes

Under the hood — without giving away the keys

A research-grade architecture, engineered end-to-end.

PRISM is not a fine-tuned public model. It is a ground-up, proprietary architecture developed by PhD-level deep-learning researchers over months of controlled experiments — leakage-audited data pipelines, walk-forward validation, adversarial benchmarking. The exact mechanisms stay in the vault; the building blocks below tell you why it works.

Market ingestion

11 markets × 4 timeframes
price · volume · flow · sessions

N× attention blocks

Multi-Head Self-Attention

Normalization · Gated FFN

Multi-Head Self-Attention

Normalization · Gated FFN

multi-resolution encoders,
weights shared across markets

Fusion core

global cross-market
attention state

48 temporal queries

every future hour decodes
its own latent state

Emission heads

OHLC path engine

Direction probability

Confidence σ

Consistency lock

×3 expert ensemble
votes on every candle

Volatility-adaptive normalization — one engine handles calm FX majors and violent crypto alike.
Deep ensembling — multiple independently trained experts vote on every candle.
Regime-robust training — randomized market blackouts teach the model to survive missing data; it still forecasts on weekends when only crypto trades.
Leakage-audited — every input is provably information-available at decision time.

Prediction power, quantified

It doesn't just predict. It knows how sure it is.

Every one of the 254 forecast candles carries its own calibrated uncertainty — tight when the model is confident, wide when the market is treacherous. And the edge is not one lucky market: it spans metals, FX majors, indices, energy and crypto.

Calibrated confidence by horizon — Self-estimated uncertainty grows honestly with the horizon — per timeframe.

Per-market error reduction — Error reduction vs the statistical baseline, market by market (4-hour candles).

What you get

One engine. Every answer a trader asks.

◔Full OHLC candles

Not a single number — complete open/high/low/close paths: 192 × M15, 48 × H1, 12 × H4, 2 × D1, all describing the same 48 hours, all exactly consistent.

▲▼Direction probability

P(up) for every candle, from the forming bar to the 2-day close — calibrated, never contradicting the drawn path.

±σCalibrated confidence

Per-candle uncertainty bands you can size positions with. Wide σ = stand back; tight σ = the model means it.

⛓11 markets, one brain

Gold, silver, EURUSD, GBPUSD, USDJPY, USDCHF, NZDUSD, S&P 500, Dow, WTI, Bitcoin — forecast jointly, not in isolation.

⟳Live every 15 minutes

Self-healing data pipeline re-anchors the forecast at every new quarter-hour candle, around the clock.

🛡Fails gracefully

Holiday? Weekend? Data outage? The engine keeps forecasting with whatever the market gives it.

Coverage

What it forecasts — exactly.

11

Supported markets

XAUUSDGold XAGUSDSilver EURUSDEuro GBPUSDPound USDJPYYen USDCHFFranc NZDUSDKiwi US500S&P 500 US30Dow Jones WTICrude Oil BTCUSDBitcoin

Metals · FX majors · US indices · Energy · Crypto — forecast jointly by one engine.

4

Supported timeframes

M1515 minutes H11 hour H44 hours D1Daily

All four views describe the same 48 hours and agree with each other exactly — the 15-minute candles aggregate perfectly into their hour, their 4-hour block, their day.

254

Future candles per forecast

M15192 candles

H148 candles

H412 candles

D12 candles

Every candle with full OHLC + direction probability + ±σ confidence, re-anchored every 15 minutes.

The market's next 48 hours,
candle by candle.