Three days in July 2026. The trial-by-trial log is RESEARCH_NOTES.md, the condensed findings are in KNOWLEDGE_BASE.md, and the literature reviews live under reports/deep_research_*.md. This file is the short version, told in order.
I started with a cross-sectional S&P signal I'd built earlier: rank ~500 stocks weekly with a gradient-boosted model, hold the extreme deciles. The IC had a t-stat of 3.35, which sounds like a finished project. It wasn't.
Cost-aware rebalancing (trade partway toward the target book instead of rebuilding it weekly, an idea from Garleanu and Pedersen) added about 0.2 Sharpe, right where the literature said it would. Then I rebuilt the index membership point-in-time and watched part of the edge turn out to be survivorship. The harshest test came last: pick the configuration using trailing data only, then measure forward. Sharpe 0.28 became 0.05. Signal real, money gone. I wrote that down and moved on.
Along the way I built new features from primary sources instead of vendors: EDGAR XBRL fundamentals, Form 4 insider filings, overnight versus intraday returns. Composite IC reached t=3.9 on a 5,594-name panel, the best statistic in the repo, and large-cap economics stayed thin anyway. Costs and crowding beat statistics.
One strategy survived every test I could throw at it. Take the 77-instrument multi-asset universe and, instead of timing each instrument, rank them against each other and trade the cross-section. Walk-forward Sharpe 0.85, clean shuffled-label control, positive in 12 of 15 years, and it passed the same config-selection gate that killed the flagship.
I replicated eight published anomalies on my own panel. Six were dead or had flipped sign. Momentum limps along. The only one still fully alive is the turn-of-month effect, published in 1968, which persists because the buying it exploits comes from payroll calendars, and payroll calendars don't read papers.
Other tests from this stretch: realized weather doesn't predict natural gas (the market trades forecasts days earlier, so whatever edge exists is in forecast errors, not outcomes). WallStreetBets attention stopped predicting anything in early 2021, which matches the academic record. Post-earnings drift in small caps is gone at every size tier I could measure, most likely because software reads every filing now.
I picked prediction markets as the focus and did the homework first: microstructure, fee models, who the institutional players are (Susquehanna has been Kalshi's market maker since 2024; DRW and Wintermute were hiring prediction-market traders in 2026), and what the academic record claims.
The famous claim is the favorite-longshot bias: favorites underpriced, longshots overpriced. I ran a betting backtest with a two-month development window and a five-month frozen out-of-sample vault, 55 registered trials in total. The frozen strategy scored below an efficient-market Monte Carlo null, at the 2.4th percentile. Worse than random betting at the same prices.
So I measured the calibration myself on 3,341 resolved markets. The bias has reversed since the 2021-25 study window: the market is now mildly overconfident, slope 0.87. And the reverse trade loses too, 21.7% per bet after spreads, because the spread is wider than the mispricing in both directions. Every directional strategy on this venue pays a toll larger than the edge. The one participant who nets positive is the market maker collecting that toll.
That conclusion set the current phase. Three mechanisms make documented money in this venue and none of them can be backtested, because they live in order books, reward programs, and other people's positions rather than in price history. So they run live, on paper, and grade themselves:
A scheduled job checks the ledgers daily and writes a digest. Nothing goes live with real money until a mechanism clears the same gates everything else faced.
About 70 registered experiments, 15 dead strategies, one survivor, a calibration result that contradicts the published record, and a set of instruments that keep collecting data whether or not anyone is watching. Also a habit that mattered more than any single result: when a metric and a profit number disagree, believe the metric.