ATLAS
OHLCV-only baseline for indices and mega-caps
Titan carrying the world — stable foundation
What ATLAS does
The minimalist. ATLAS uses only the five raw OHLCV channels — no indicators. Smallest of the Five (2.4M params), highest win rate at symmetric 1:1 R/R, and the most conservative of our models.
Broad-market trend following, conservative R/R
How it works
ATLAS is a factorised channel × time attention transformer. It looks at a 120-bar context window of 5 input channels, and predicts the full distribution of the next 5 bars.
Forward pass: input OHLCV + indicator channels → patch-embedded → alternating space/time attention layers → pooled to a single embedding → MLP head predicts 7 quantile levels per horizon step.
Input channels
Every bar is z-score normalised per-window using context-only statistics, so the model sees relative moves rather than absolute prices.
Output
Data & training
Universe
120 symbols
indices_and_megacaps
Symbols are the liquid, tradeable instruments that form ATLAS's training population. Each prediction cycle runs forward-tests on the same universe so performance numbers aren't cherry-picked.
Why this training window
We deliberately exclude pre-2010 data. Pre-decimalisation (fractions until April 2001), pre-HFT regime (≤2007), and the 2007-2009 credit crisis reflect a market structure that no longer exists. At 2.4 M parameters, burning capacity on that regime is noise competing for weights with the current market — López de Prado (2018) identifies regime non-stationarity as the primary failure mode of financial ML.
Sensitivity: a pilot trained on 2003-2022 ran validation loss 10-12% worse than the 2010-onwards run.
Live performance
Every prediction ATLAS makes is logged and barrier-evaluated in real time. These numbers reflect closed trades only and update continuously.
Past performance does not guarantee future results. Forward-test outcomes reflect 1:3 R/R barrier simulation; live outcomes may differ due to slippage, spreads, and liquidity.
Honest limitations
Bull-regime test exposure
The 15-week out-of-sample window was a bull-biased period. Model performance in bear regimes (2022-Q4 style) is unverified and likely materially lower.
Transaction costs not modelled
Headline returns exclude round-trip fees (~0.05-0.30% on these instruments). Real-world returns will be lower; headline numbers are upper bounds.
Same-symbol train/test
The model was trained and tested on the same symbol universe (time-split). We validate temporal generalisation, not universe generalisation to unseen tickers.
What we're confident about
Directional skill above random chance (symmetric 1:1 R/R win rate is 51.8%) and reproducible training: we've retrained from scratch 3+ times with identical hyperparameters within ±1% val loss.
Read the full paper
Publication-depth technical report: data cleaning, architecture deep-dive, training dynamics, barrier-simulation methodology, full-window results, sensitivity analyses, and honest-limitations discussion. ~7,000 words.
Compare across the Five
Every model in the Zirdle Five tackles a different trading regime. Head back to the overview to see how they stack up on win-rate, universe size, and out-of-sample returns.