ARA.AI

Ara.AI — Stock Ranking Models

This repository hosts two generations of model. v8 is current; the v7 transformer checkpoints are kept for reproducibility only.

Source: https://github.com/MeridianAlgo/AraAI · Design notes: docs/ARA_V8.md

What v8 predicts

For each stock on each day, the next-day return relative to the universe average that day — not the absolute return. The output is a ranking; the intended use is a dollar-neutral long/short book (long the top names, short the bottom ones).

This matters for reading the numbers: a score of +0.004 means "expected to beat the universe average by about 40 bp tomorrow", not "expected to rise 0.4%". A market-neutral model does not forecast market direction and should not be compared against an always-up baseline.

Architecture

• HistGradientBoostingRegressor × 3 seeds, averaged (400 iterations, 15 leaf nodes, lr 0.03, L2 1.0, max_features 0.7, no early stopping)
• ~40 scale-invariant daily features per (symbol, date): multi-horizon returns, realized vol and vol-normalized shocks, SMA distances, intraday range/gap structure, ATR, Wilder RSI, volume z-scores, 52-week distances, calendar
• 8 of those additionally as cross-sectional within-day percentile ranks
• Target: cross-sectionally demeaned forward return, winsorized at 4 per-day σ
• No price or volume levels anywhere (they encode symbol identity, not signal)

Performance

Expanding-window walk-forward, 4 folds over 2025-06-02 → 2026-08-06, retrained from scratch before each fold with a 1-day embargo between train and test. 99 symbols, median 70 names per day, 297 test days.

Positive in all four folds and stable across seed groups (IC +0.0204 to +0.0209, t 2.25–2.30). t = 2.26 clears the conventional bar only just, on a single window, and one fold (+0.0405) carries much of the average while the other three sit near +0.013 with t < 1 individually. The long/short figures are pre-cost and would not survive daily turnover on a five-name-per-side book at retail commissions. Treat this as a small measured edge, not a trading system.

v7, for comparison

Trained on data before 2025-06-01 and evaluated on the year after:

v7 had no edge and its magnitude forecast sat exactly on the zero-prediction floor. Any higher figure in older documentation came from a CI checkpoint that had trained through its own evaluation window.

Limitations

• Stocks only. 22 FX pairs is too thin a cross-section to rank, and the source FX daily bars leak next-day information through day-t high/low.
• 99-name universe. Median 70 tradeable names per day and 5 per side means few independent bets, which is why the long/short P&L is much noisier than the rank IC.
• Daily close-to-close only. No intraday, no multi-day horizon.
• Pre-cost. No transaction costs, slippage, borrow, or capacity modeling.
• Survivorship. The symbol list is today's large caps; delisted names are absent from the history.

Usage

import joblib

p = joblib.load("ara_v8_stocks.joblib")   # {"models": [...], "features": [...], "meta": {...}}
# Build features with `ara.make_dataset` from the repo, then:
# scores = np.mean([m.predict(X) for m in p["models"]], axis=0) / 100.0

The repository provides the whole path:

pip install -r requirements.txt
python -m ara predict --db-file training.db --model-path models/ara_v8_stocks.joblib

Disclaimer

Research and educational use only. Not financial advice. Past performance does not guarantee future results, and the figures above are pre-cost and not statistically significant. Never trade with money you cannot afford to lose.