Open30 Research Report

This report summarizes the research problem, methodology, walk-forward validation design, tracked results, limitations, and next experiments for the Open30 strategy research stack.

Summary

Open30 studies whether opening-session price action, prior daily context, market alignment, and news sentiment can identify a favorable 30-minute trade candidate after the market open. The current research stack builds candidate rows keyed by (date, ticker, side), labels take-profit, stop-loss, and time exit outcomes, then evaluates model decisions with rolling walk-forward training.

The strongest tracked full-history single-head branch is v1_3. It traded 1,865 days, averaged 56.82 bps per selected trade, and had a 1.169 profit factor in the tracked artifact. The improvement audit finds that this edge is execution-sensitive and should be hardened with realized cost modeling, delayed-fill tests, point-in-time universe construction, and stronger calibration.

Problem Statement

The strategy asks this: shortly after the open, is there one liquid large-cap equity candidate with enough expected return to justify a short-horizon trade through the first 30 minutes?

The research version evaluates candidates using historical minute bars, daily features, market context, and sentiment features. It does not claim a deployable trading system by itself.

Data And Universe

The pipeline expects:

• 1-minute bars at data/raw/candles_1m.parquet.
• Daily news files under data/raw/news_daily/.
• A tracked top-25 universe file at data/interim/canonical/universe_top25.json.
• Generated features, instances, labels, and final dataset under data/processed/.

The tracked result metadata records 195,166 dataset rows and 3,983 trading days for the main full-history runs. The public pipeline config starts ingestion on 2010-02-02; feature warm-up means the main tracked backtest rows begin on 2011-10-06. The fixed present-day top-25 universe is a known risk because it can overstate historical performance through survivorship and concentration effects.

The assembled dataset_open30m.parquet artifact is also available on Hugging Face for readers who want to inspect or reuse the feature table without rerunning ingestion and feature generation: mospira/open30-equity-features.

Strategy Definition

The core strategy invariants are:

• Candidate rows are built per (date, ticker, side).
• Entry is the 09:31 open.
• Take-profit and stop-loss scanning runs from 09:31 through 09:59.
• The time-exit proxy is the 09:59 close.
• Class mapping is 0=SL, 1=TP, 2=TIME.
• Ambiguous rows are converted to worst case before simulation and scoring.
• Current canonical execution is long-only at decision time.
• At most one trade is selected per day in the canonical backtest.

The architecture manifests in architectures/ control reward multiples, selection mode, training window, retrain cadence, embargo, calibration, stop distance, EV thresholds, and position sizing.

Methodology

The pipeline builds features in stages:

1. Fetch minute bars and news sentiment.
2. Canonicalize sentiment to daily ticker-level records.
3. Build daily, open-window, market-context, and sentiment features.
4. Assemble feature rows keyed by (date, ticker).
5. Build candidate trade instances keyed by (date, ticker, side).
6. Generate labels for reward multiples.
7. Assemble data/processed/dataset_open30m.parquet.

The backtest uses rolling walk-forward retraining with a 730-day lookback, 30-day retrain step, and 1-day embargo in the tracked full-history artifacts. Training currently uses long-side rows unless an experiment deliberately changes that policy. Probability calibration uses isotonic regression.

Architecture Versions

Results

Single-Head v1.x Branch

Account curves are sensitive to starting capital and margin policy. The table therefore emphasizes trade-level metrics. open30_v1 also used a different starting capital from v1_1 through v1_3, so account return should not be read as a clean architecture comparison.

The metric comparison highlights the central v1.x tradeoff: the threshold variants select fewer trades, but the selected-trade return and profit factor improve relative to the raw baseline.

The normalized equity chart puts the v1.x variants on the same starting value. It uses a log scale because account values compound across a wide range.

The threshold variants reduced trade count while improving mean selected-trade return. Drawdowns remain large, so this is not yet a robust production claim.

Multi-Head v2

The multi-head expected-return selector did not beat the stronger single-head branch in the tracked artifacts.

The m=1.5 head had the best mean trade return in this v2 run, while the Best EV selector increased trade count but had the weakest drawdown profile.

Stop-Distance Sweep

The stop-distance sweep covers a shorter recent period, from 2024-11-21 through 2026-05-22. It should be read as an experiment about stop geometry, not as a full-history replacement for the v1.x branch.

The recent stop sweep shows unstable stop-distance behavior: k=0.2 had the highest account return over this window, while k=0.3 had the highest profit factor. The dynamic selector did not dominate the static settings.

Improvement Audit

The improvement audit concluded that the long-only m=1.5 edge is real enough to continue researching, but very sensitive to execution assumptions. The most important tested improvements were:

• Use a fixed, cost-aware EV floor near 0.10 for the long strategy.
• Train a separate short-only model with a stricter EV floor near 0.25.
• Deduct realistic costs and slippage from realized backtest PnL.
• Replace the fixed present-day top-25 universe with point-in-time membership.
• Explore staggered retrain consensus before heavier model families.

Reproducibility

The public export did not rerun pipelines. To reproduce results locally, either download dataset_open30m.parquet from mospira/open30-equity-features and place it at data/processed/dataset_open30m.parquet, or rebuild it with the full pipeline:

python run_pipeline.py --architecture architectures/open30_v2.yaml
python run_backtest.py --architecture architectures/open30_v2.yaml
python run_backtest.py --architecture architectures/v1_3.yaml
python run_v1_stop_distance_sweep.py

Generated outputs are written to reports/. Curated outputs in results/ are reviewed snapshots copied from those generated reports.

Limitations

These are the main blockers before treating results as production-ready:

• Fixed present-day top-25 universe may introduce survivorship bias.
• Realized PnL does not yet fully model costs, slippage, and delayed fills.
• Labels assume the exact 09:31 open.
• Calibration and threshold tuning need stricter cross-fitting.
• Dynamic threshold objectives can favor low trade counts rather than daily net return.
• Retrain-calendar phase sensitivity can materially change outcomes.
• Ticker and head concentration remain meaningful.

Next Research Direction

The next high-value work is to make the simulation more execution-aware, rebuild the universe point-in-time, and then retest the simpler m=1.5 classifier/EV design with fixed cost-aware thresholds before promoting more complex model families.