No trading idea earned real money. Every test used simulated (paper) money or historical data, and the research is now paused.

WhyAn early study found spot crypto prices move before Kalshi's contract prices, suggesting the market might be slow and beatable by a model.

DataKalshi hourly Bitcoin/Ether contracts (KXBTC, KXBTCD, KXETH, KXETHD) with 1-minute price history plus spot crypto prices (from FMP); pinned research dataset spanning roughly 2026-05-10 to 2026-07-08.

MethodAn AI agent repeatedly edited a single model file and was scored by a frozen, tamper-proof scorer (a Sharpe-style score across 5 data slices). A final 'sealed' test set was held back. Before spending it, the winning model was run as a zero-money paper trader on live markets to see if its claimed edge showed up.

Cost$0 real money; about 2 days of paper trading to falsify

WhyThe v1 post-mortem blamed a scorer that let the model peek at future data and a single-regime dataset. v2 fixed both, so a model that scored well here should be more trustworthy.

DataKalshi hourly Bitcoin/Ether contracts on the full archive (about March to July 2026, roughly 2.2 times v1's history), with spot prices; about 41,000 simulated trades in the winning model's research score.

MethodWalk-forward testing: each test period is scored only with a model trained on earlier data. The model also had to be better calibrated than the market on the exact trades it would take. An AI agent ran 572 trials. The sealed final test would only be spent if a live paper trader confirmed the edge over about two weeks.

Cost$0 real money

WhyIf Kalshi lags spot, a fast trader could act in the gap. It could also explain why the 5-minute paper trader underperformed.

DataOur 10-second quote recordings (spot quotes from FMP plus Kalshi best bid/ask), 2026-07-16 to 07-20, 5 days; for example 600 KXBTCD markets.

MethodCorrelated each spot move with Kalshi price changes at delays from -60 to +300 seconds. A clean result shows nothing at negative delays (Kalshi never leads) and a peak at some positive delay.

CostNo direct cost recorded.

WhyPaper trading is the cheapest honest check before risking money. The gap between predicted and realized profit also measures how much a trader loses to better-informed counterparties (adverse selection).

DataOur own paper-trading ledger (prediction/paper/ledger.jsonl): 222,631 settled one-contract v2 trades on Kalshi crypto contracts, 2026-07-17 to 2026-09-23.

MethodEvery 5 minutes the model picked contracts it thought were mispriced and logged a pretend order, filled at the next real quote. At settlement each trade's result was compared with the model's expected profit at the time of the decision. A side study checked whether stale data explained the losses.

Cost$0 real money; ran 2026-07-17 to 2026-09-23

WhyTaking liquidity lost about 6 cents a trade. Spot leads Kalshi by about 20 s, maker fees are $0, and H7 shows when not to post, so a fast maker bot was the remaining candidate.

Data1,110,146 event-driven decision moments from our 5-second quote recordings, 2026-07-21 to 08-26 (28 research days, 1 embargo day, 8 sealed days), with a frozen fill simulator that only counts an order filled if the price trades strictly through it.

MethodPre-registered design on 07-21, before the data existed. The scorer requires positive skill and better calibration than the market on filled trades in each of six time slices; the starting score is 0 by design. An AI agent then ran about 340 experiments in one 12-hour window.

Cost$0 money; one 12-hour run on the local LLM

WhyBetting markets often overprice longshots (the favourite-longshot bias). If Kalshi crypto shows a stable pattern, a model could correct prices toward reality.

DataSettled Kalshi crypto markets with 1-minute price history: about 35,000 liquid markets 2026-01-01 to 07-25, then a fresh strip after 07-25; weekly cumulative readings to 09-20 (about 93,000 market readings per horizon).

MethodGrouped contracts by price (0-10 cents, 10-20 cents, ...) at 30, 15 and 5 minutes before close, and compared the win rate with the price. The re-test used confidence bands that account for related strikes and checked whether the pattern held on fresh data.

CostNo direct cost recorded.

WhyBitcoin is thought to lead price discovery in crypto. A cross-asset volatility clue could be a cheap feature for a crypto model.

DataSpot Bitcoin and Ether quotes from our seconds-level quote recordings (quotes from FMP): 11 days at first (to 07-26), a fresh 31-day window 07-27 to 08-26, then weekly re-reads to 09-20 (66 days).

MethodMeasured Bitcoin's choppiness with the 'Roll' estimator after removing what Ether's own state explains, then correlated it with Ether's future volatility change. A shuffled-days placebo checks it is not a time-of-day effect. A frozen re-test on fresh data followed.

CostNo direct cost recorded.

WhyThe paper traders were losing about 6 cents a trade to better-informed counterparties. A signal for when not to post orders could reduce that loss for a maker strategy.

DataOur 5-second quote recordings, 2026-07-16 to 07-30 (15 days, 13,429 markets, about 781,000 usable rows); fills simulated from real quote movements.

MethodPre-registered test: a book-activity score was computed for each moment and split into four buckets using cutpoints fixed on early data. On later held-out data we measured how much each simulated fill lost 60 seconds later. Pass required at least 1 cent between calm and hot books plus a randomization test.

CostNo direct cost recorded.

WhyThe same spot-leads-Kalshi pattern had appeared in crypto and stock indexes; testing a rates market checks whether it generalises.

DataKalshi KXUST10AD contracts (series began 2026-07-08) with 1-minute prices, and minute prices of a Treasury bond fund (from FMP), 2026-07-01 to 08-27: 360+ markets, about 45,000 aligned minutes.

MethodPre-registered lead/lag test that needed a complete ladder of strike prices quoted at the same minute for at least 30 minutes in a row on at least 10 days. Eligibility gates were checked before any result was computed.

Cost20 FMP + 240 Kalshi candle calls for the re-collection; $0

WhyA third asset class (after crypto and stock indexes) would show whether 'spot leads Kalshi' is a general pattern.

DataKalshi Brent contracts (KXBRENTD, KXBRENTW) with 1-minute prices and Brent spot minute prices (from FMP), 2026-07-01 to 07-30: 29 days, 549 usable markets.

MethodPre-registered test with a randomization null and an economic floor: the lead had to be worth at least the 3-cent cost of posting (2-cent spread plus 1 cent).

CostNo direct cost recorded.

WhyAnother asset class for the spot-leads-Kalshi pattern, using a currency the forex program already covered.

DataKalshi EURUSD series (daily KXEURUSD, hourly KXEURUSDH, weekly KXEURUSDW) and EURUSD minute prices (from FMP), 2026-07-08 to 08-06; 330 markets kept.

MethodA coverage-only check (no results computed) to see if the pre-registered eligibility rules could be met before running the actual test.

Cost30 FMP + 330 Kalshi candle calls; $0

WhyCopying politicians' trades is a popular idea, the data was already included in the FMP subscription, and a paper-trading stock account (paper broker) had just become available as a place to test it. The coverage check showed filings arrive a median 27 days after the trade, so the only testable question was whether the filing itself still moves the stock.

DataCongressional trade disclosures from FMP (Senate and House feeds; 39,164 filings found in the coverage probe), limited to March 2023 to August 2026 when both chambers are covered, stocks only, de-duplicated; daily stock and SPY prices from FMP. After filters: 14,745 stock-day events; 11,821 research-window events (5,270 buy events, 6,551 sell events).

MethodAll rules, tests, a cost floor and pass/fail criteria were written down and frozen before any return was computed (pre-registration). The simulated trade buys at the first market open after the filing date and measures return against the S&P 500 fund (SPY) after 1, 5 and 21 trading days. A 'placebo' ran the same measurement on ordinary days for the same stocks, and the most recent six months (March-August 2026) were locked away as an untouched final test to be used only if the main test passed. The analysis was run exactly once.

CostAbout two days of work and ~2,200 FMP data calls; $0 of real money.

WhyInsider purchases are among the best-documented signals in the academic literature, and unlike congressional filings they become public within two business days. It reused the machinery just built for the congressional study.

DataInsider purchase filings from FMP (open-market purchases only, Form 4, January 2023 to August 2026) and daily prices from FMP. Only 'tradable' stocks were kept (price at least $5 and at least $1M of typical daily trading). 11,752 tradable research-window events; 18,670 untradable events excluded.

MethodSame approach as the congressional study: rules, five tests, a cost floor and a placebo were frozen before any results. The simulated trade buys at the next open after the filing date and compares to SPY over 1, 5 and 21 trading days. The last six months were kept locked as a final test, and the analysis was run exactly once.

CostOne working session across two days and 5,091 FMP data calls; $0 of real money.

WhyTrading fees on Kalshi were a concern, and Polymarket US (a newly regulated US venue) charged lower taker fees and even paid makers. A second venue would only matter if its order books were actually usable.

DataPublic order-book snapshots from both venues every 5 minutes, 19 clean days (2026-08-08 to 08-26): 521,778 usable records across 245 matched pairs (MLB and NFL game winners plus 4 bitcoin price markets). No accounts were needed.

MethodA collector matched identical markets across the two venues and recorded both order books side by side. The report compared median spreads and the number of contracts available within 1 and 5 cents of the middle price. This was a measurement, not a trading test.

CostAbout three weeks of automatic collection (about 1.7 GB of order-book data kept); $0.

WhyThe favorite-longshot bias is a well-known betting pattern, the test was free (Kalshi's public archive of settled games), and the same technique had already been used on crypto markets. It was the recommended next program after the filings studies.

DataKalshi's public archive of settled game-winner markets: 29,662 markets / 14,831 games across 7 leagues (MLB, NBA, NHL, WNBA, NFL, college football, college basketball), April 2025 to September 2026. One side (home team) per game, priced 5 hours before the market closed. 12,281 games in the research window.

MethodFirst a check that the data existed without looking at any outcome. Then rules, five tests, a minimum-size floor and a final locked test block (July-September 2026, 1,177 games) were frozen before the analysis. The single run compared each price band with how often those teams actually won.

CostTwo days and ~16,000 free Kalshi API calls; $0.

WhySports was the area of interest after the crypto results, and Polymarket US had emerged as a possible cheaper second venue for sports.

DataDesk research plus the Kalshi vs Polymarket US depth data; no new experiments ran from the brief itself.

MethodThe brief ranked five ideas (H-Sports-1 to 5) by cost and evidence, gating paid data behind free tests first. H-Sports-5 (price fairness) was run separately as a full pre-registered study.

CostNo direct cost recorded.

WhyKalshi lists markets on economic releases, and the FMP subscription carried an economic calendar with actual numbers. A strategy that reads the number and trades within seconds only works if the data arrives within seconds.

DataThe script's own timing measurements: FMP's economic calendar checked every 2 seconds around 11 weekly jobless-claims releases (2026-07-16 to 2026-09-24). Only the delay in seconds is reported here.

MethodA go/no-go measurement before building anything, with the kill condition set in advance: if FMP lags the release by minutes, the strategy dies. After the first reading a weekly Thursday job kept measuring to see if it was representative.

CostNo extra cost beyond the existing FMP subscription; a small weekly job.

WhyThe subscription was heavily under-used (about 0.06% of its call limit in July), so its data was surveyed for more uses.

DataLive probes of the FMP subscription (July 2026) and a literature sweep.

MethodIdeas were ranked by value and effort. The top three were implemented or measured (fast quote polling, the index lead-lag study, and the economic-release timing test). The rest were deliberately parked.

CostNo direct cost recorded.

WhyIt is a low-risk idea with a clear score (shares accumulated per dollar), and it would have been the project's first live link to a real brokerage.

Data10.5 years of hourly SPY price bars (2016-2026) from FMP, plus published research on dollar-cost averaging.

MethodBefore building anything, the brief measured the best possible result with perfect hindsight (the ceiling for any model) and tested simple rules such as buy-the-dip against plain even buying.

CostNo direct cost recorded.

WhyCurrencies visibly react to data surprises, and the project had a news/surprise feature pipeline ready; it was the first, most natural idea to test. It was also run across the other USD majors (EUR/USD etc.) before the USD/JPY sealed test.

DataDukascopy 5-minute bid/ask prices for USD/JPY (about 5.4 years, 2021 to mid-2026), plus an economic calendar with forecasts, central-bank rate decisions and market rates (US/Japan 2-year yields, VIX) — calendar and rates from FMP (no raw rows shown here).

MethodAn AI coding agent running on a local model repeatedly edited one strategy file inside a locked 'sandbox' and was scored by a frozen scorer on the research years only, charging real bid/ask spreads on 5-minute bars. Every attempt was counted so the final score could be 'deflated' (penalised for how many tries it took). The last ~year of data (2025-04 to 2026-06) was sealed and opened exactly once, by the owner, at the end; to pass, the deflated sealed score (test_dsr) had to be at least 0.05 and most sealed windows and calendar years had to be positive.

CostAbout $0 in trading (never deployed); one ~8-hour agent run on local hardware.

WhyTo either find a real edge or prove that the search process is honest — i.e. that it kills ideas that only look good in hindsight.

DataUSD/JPY (and earlier EUR/USD and other majors) 5-minute Dukascopy prices, ~5.4 years, plus calendar, rate and market data.

MethodFour evidence-ranked theses, each searched by an AI agent under a frozen scorer, each given exactly one sealed test opened by the owner. A stopping rule was pre-registered before the last thesis.

CostAbout $0 in trading losses; roughly nine days of machine time (2026-06-26 to 07-04).

WhyA literature survey ranked intraday time-of-day patterns as the only family with repeated, tradable evidence on the major pairs that could be tested with the 5-minute data already on hand.

DataDukascopy 5-minute bid/ask prices for USD/JPY, about 5.4 years (2021 to mid-2026), with real per-bar spreads.

MethodAn AI coding agent running on a local model repeatedly edited one strategy file inside a locked 'sandbox' and was scored by a frozen scorer on the research years only, charging real bid/ask spreads on 5-minute bars. Every attempt was counted so the final score could be 'deflated' (penalised for how many tries it took). The last ~year of data (2025-04 to 2026-06) was sealed and opened exactly once, by the owner, at the end; to pass, the deflated sealed score (test_dsr) had to be at least 0.05 and most sealed windows and calendar years had to be positive.

CostAbout $0 in trading; one agent run on local hardware.

WhyCarry is the most-replicated pattern in currency research, and 2021-2026 contained every carry regime (zero gap, fastest widening in decades, the August 2024 carry crash, narrowing) — a rich test.

DataDukascopy 5-minute bid/ask prices for USD/JPY (about 5.4 years, 2021 to mid-2026), plus an economic calendar with forecasts, central-bank rate decisions and market rates (US/Japan 2-year yields, VIX) — calendar and rates from FMP (no raw rows shown here). Policy-rate series were derived from central-bank decision events in the calendar.

MethodAn AI coding agent running on a local model repeatedly edited one strategy file inside a locked 'sandbox' and was scored by a frozen scorer on the research years only, charging real bid/ask spreads on 5-minute bars. Every attempt was counted so the final score could be 'deflated' (penalised for how many tries it took). The last ~year of data (2025-04 to 2026-06) was sealed and opened exactly once, by the owner, at the end; to pass, the deflated sealed score (test_dsr) had to be at least 0.05 and most sealed windows and calendar years had to be positive.

CostAbout $0 in trading; one agent run.

WhyThesis 3 died at a regime turn; market-implied expectations were designed specifically to anticipate such turns. A stopping rule was declared beforehand: if this failed, Program 1 would end.

DataDukascopy 5-minute bid/ask prices for USD/JPY (about 5.4 years, 2021 to mid-2026), plus an economic calendar with forecasts, central-bank rate decisions and market rates (US/Japan 2-year yields, VIX) — calendar and rates from FMP (no raw rows shown here).

MethodAn AI coding agent running on a local model repeatedly edited one strategy file inside a locked 'sandbox' and was scored by a frozen scorer on the research years only, charging real bid/ask spreads on 5-minute bars. Every attempt was counted so the final score could be 'deflated' (penalised for how many tries it took). The last ~year of data (2025-04 to 2026-06) was sealed and opened exactly once, by the owner, at the end; to pass, the deflated sealed score (test_dsr) had to be at least 0.05 and most sealed windows and calendar years had to be positive. A pre-registered stopping rule (declared 2026-07-04 before any result) made this the last single-pair thesis.

CostAbout $0 in trading; one ~6-hour agent run.

WhyProgram 1's diagnosis was too few independent regime bets. Cross-sectional carry/momentum portfolios are where the surviving academic evidence lives, and 14 years of data (2012-2026) covers many more regimes.

DataDukascopy 5-minute prices for EUR/USD, USD/JPY, GBP/USD, AUD/USD, USD/CAD from 2012 to 2026 (14 years), with policy-rate differences from the economic calendar (available from 2015) and VIX. Calendar/VIX from FMP (no raw rows).

MethodSame locked-scorer agent search, now scoring a weekly-rebalanced basket with each pair's real spread. The last ~20% of history (2023-2026, including the August 2024 carry crash) was sealed and opened once by the owner. Program rules pre-registered: at most two iterations before a full human review.

CostAbout $0 in trading; one 6-hour agent run.

WhyA forward clock cannot be contaminated by the search. It is the cleanest check on whether the +0.37 sealed result was real.

DataNightly Dukascopy price catch-up for the five pairs; our own append-only ledger (forex/paper/ledger.jsonl) of weights and daily net returns, 57 forward trading days (2026-07-07 to 2026-09-24).

MethodTwo strategies were frozen on 2026-07-06. 'trend_basket' (control): each week, for each of the five currencies vs the dollar, hold +20% if it rose over the last 100 days and -20% if it fell. 'xsect_best_16e37f8' (research winner): each week rank the five currencies by a mix of interest-rate advantage (scaled by how jumpy it is) and smoothed recent trend; put 50% long the top one and 50% short the bottom one; go flat if VIX >= 30; trim 20% when markets are unusually volatile. Returns are net of real spreads and slippage. The first read was pre-registered for +90 days (2026-10-04); it was taken early for reasons unrelated to results.

Cost$0 capital at risk; nightly cron job on existing hardware.

WhyProgram 1 concluded text/perception features were one of the few untested directions; the news-sentiment tooling had been built but never fed a real corpus.

Data6,805 full-text forex news articles (2024-06 to 2026-07, from FMP, accruing nightly until retirement) and 11 years of GDELT global media tone (2015-02 to 2026-07, 9,434 day-currency rows, via BigQuery at $0). The NYT feed was retired 2026-09-03.

MethodHeadlines were scored by FinBERT (a finance-tuned language model) with uncertain cases escalated to a local LLM. A 300-headline study compared the two; then 3,261 LLM labels 'anchored' to a specific currency were used to fine-tune FinBERT, tested on 610 later (time-held-out) items. A protocol was written for letting trained models into the locked research harness without leaking future information (models trained only on earlier folds, fitted once for the sealed test).

CostGDELT $0; FMP subscription (since cancelled) shared with other uses; GPU time on the local machine.

WhyKalshi and FMP only provide minute-level history. Measuring second-level effects (the ~20 s lag, which orders get filled) required recording it live.

DataSpot quotes for BTC, ETH, SPY, QQQ and VIX (from FMP) plus Kalshi top-of-book for 6 series; 72 daily files 2026-07-16 to 09-25, about 926 MB.

MethodA small always-on recorder, restarted every minute by cron if it stopped, wrote only changed values. The 5-second rate was chosen to put several samples inside the 10-40 s repricing window.

CostPart of the FMP subscription; no extra cost

WhyLeadership between markets can change when volatility is high, so a live VIX reading was a cheap candidate conditioning input.

DataVIX from FMP, added to the daily bar collector and to the 5-second quote recorder from 2026-07-16.

MethodParked as a standalone strategy. It was later tried inside v3's search as a gate (trade only in certain VIX conditions).

CostNo direct cost recorded.

WhyAn agent can try hundreds of ideas overnight for free on local hardware. If the scoring is honest and the final exam is kept hidden, the machine can do the tedious search while humans keep only the decisions.

DataForex: Dukascopy 5-minute USD/JPY and EUR/USD bars with real spreads (2021-01 to 2026-06, about 406,000 bars), economic-calendar surprises, rates and VIX; the portfolio run used 5 major pairs from 2012. Prediction: Kalshi BTC/ETH price-range contracts on a 5-minute grid (about 67,000 rows for v1, the full archive from March 2026 for v2) and 1.1 million quote-level rows for v3 (2026-07-21 to 08-26).

MethodEach project has sandboxes (one folder per idea). Inside, the agent may edit only one file (strategy.py or train.py). The 'ruler' (prepare.py, score.py, the dataset and its checksum) is frozen: read-only and hash-pinned, so the agent cannot win by moving the goalposts. Every attempt is scored on a validation slice of past data. If the score beats the best so far, the agent keeps it with a git commit. If not, it reverts. That is 'keep/revert'. The newest slice of data is the sealed test. It stays locked until the search ends, and then a human opens it exactly once, with the score discounted for how many ideas were tried. Pre-registration means the pass/fail rules and stopping rules were written down before any result was seen. A supervisor ('outer loop') runs every 30 minutes as a headless Claude session on cron. It checks that the run is alive and intact, fires tripwires on suspicious jumps, can pause a run, and texts the human over WhatsApp. It never opens the sealed test.

Cost$0 in trading losses; local hardware and a Claude subscription for the supervisor