The honest quant journey.
We ran a real $5K trading exploration across six venues — prediction markets, equities, on-chain sports — under a paper-mode kill switch. The trading itself returns single digits. The methodology, the contamination audits, and the validation stack we built along the way are reproducible — and we publish them here in full. Most quant content is hype. This is the receipts.
Directional alpha is unlearnable on calibrated markets.
We proved it with inverse RL on a hindsight-optimal policy. Predicting the winning side 80% of the time still loses money once fees and slippage hit.
5 of 6 backtest corpora were contaminated.
Post-resolution leakage, full-sample fits, misaligned snapshots. We had to rebuild from scratch and ship a contamination test suite as a public dataset product.
Winners harvest mechanical structure.
Across 43 purged-CV-robust strategies, every winner maps to a mechanical quirk: settlement convergence, maker rebates, oracle latency, lead-lag. Pure directional bets: zero.
Pillar posts
The five long-form articles that anchor the series. Each one is the public version of an internal research engagement.
Directional Alpha Is Unlearnable on Calibrated Markets
We trained an inverse reinforcement learning agent on a hindsight-optimal policy across 45,000 prediction markets. It predicted the winning side 80% of the time — and would have lost money. The math of why, and what to do about it.
How 5 of 6 Backtest Corpora Were Contaminated
A genuine-corpus audit across six prediction-market and equity universes. The contamination patterns are common, structural, and almost never detected: post-resolution price embedded in features, full-sample isotonic fits, snapshot-misaligned price histories. Five concrete defenses.
Winners Harvest Mechanical Structure, Not Predictions
After dissecting 43 purged-CV-robust strategies across four venues, every winning strategy mapped to a mechanical quirk — settlement-window convergence, maker rebates, oracle latency, cross-asset lead-lag. Pure directional bets won zero seats. The edge-class distribution and what it means for strategy design.
The 26-Method Validation Battery Most Quants Skip
A five-family validation stack: statistical robustness, look-ahead audits, execution realism, regime stratification, and forward-paper. Out of 43 strategies that passed our purged-CV gate, zero passed the full battery. The methods, the order, and why each one matters.
Per-Venue Edge Reality on $5,000
A US-legal venue landscape audit: Polymarket-US, Kalshi, IBKR ForecastTrader, Coinbase Derivatives, Limitless. Per-venue fee structures, maker-rebate economics, and an honest portfolio allocation that returns +6–14% base / +15–25% optimistic on a $5K experimental seed. No hype.
From research → paid product
Three things the research produces that customers buy. All powered by the same engine that runs the posts above.
Validation-as-a-Service
Submit a strategy spec or backtest. We run the full 26-method battery + purged-CV + contamination tests. Pass/fail report in 7 days.
Reproducibility Datasets
Genuine corpora — Kalshi 305K markets, Limitless 36K markets, Polymarket YES-only — packaged with the contamination-test suite that proves provenance.
White Paper
"Building Honest Quant Validation From First Principles." 36 pages. The full 5-family stack + corpus-contamination audit method. Email-gated, white-labelable.
Who's behind this
Empire Research is the public methodology log of Elite AI Empire — a portfolio of focused AI products spanning consumer SaaS, research, and a small internal trading exploration. Every post here is sourced from an internal engagement ($5K experimental capital, paper-mode preserved, no public investment vehicle). We share what worked, what didn't, and the methodology in detail because the methodology is what's reproducible — the trade results are not.
If you have a strategy you'd like validated under the same battery, see Validation-as-a-Service. If you want the raw corpora to reproduce results yourself, see the datasets. If you just want the methodology in one PDF, grab the white paper.