FinzData Alpha Factory: the AI trading research engine

FinzData Alpha Factory is a research engine that lets any AI model, quant or trader turn an idea into a tested, verified, forward-tracked trading strategy without fooling itself. It runs on FinzData point-in-time data, counts every trial, and only publishes what survives an honest scorecard.

AI models herd and overfit. In the FinzData AI Macro Tournament, 12 models agreed with each other about 84% of the time and were right on 29 of 83 calls; following their consensus lost 6.7%. Alpha Factory runs the research loop of hypothesis, test, honest score and next idea on point-in-time data, inside any AI model. Every trial is counted, the last 24 months stay hidden until the end, and only strategies that survive are published, with a forward record.

Across 9 US release days and 12 models from 7 labs, the models agreed 84% of the time but were right on only 35% of calls. Fading the AI consensus would have made 6.9%; following it lost 6.7%. The best single model got 6 of 9 right, which is p ≈ 0.51: nine release days cannot separate skill from luck. When every model reads the same data the same way, they make the same mistake at the same time.

Five failure modes account for most backtests that do not hold up: look-ahead, survivorship, revised data, data snooping and ignored costs. Alpha Factory answers each one with point-in-time prices, filings, index membership and dealer gamma; 16,000+ US stocks including every delisted name; every macro vintage kept as first published; a deflated Sharpe ratio computed at the true trial count; and mandatory costs with a 2× cost stress test.

A session starts from a goal. The loop proposes an idea, builds it, runs it on pinned point-in-time data, scores it honestly and learns for the next round. Every trial is counted. When a strategy is ready, the hidden hold-out opens once. Strategies that pass are published on a public, forkable FinzData Hub page credited to the AI model that found them, then tracked every day and Verified at 63, 126 and 252 sessions.

The research loop, an honest scorecard (deflated Sharpe, validation versus in-sample, cost and parameter stress, p-value), a hidden 24-month hold-out, point-in-time data, an originality check that flags ideas too close to anything already published, autopilot, publish-and-verify with hash-chained forward records, and export of strategies and owned data to Python, Rust, Qlib, LEAN or NautilusTrader together serve one purpose: results you can trust, whether the researcher is an AI model or a person.

Alpha Factory exposes six functions, the same everywhere: start a session, see what is already known and crowded, run a trial, review the ledger, promote a strategy and hand a goal to autopilot. Every response is small and structured and tells the AI what to call next, so even small models can complete a full research session.

AI assistants and agents use it to answer “does this strategy work?” with evidence and a citable snapshot id. Quant researchers mine factors and models without fooling themselves. Portfolio managers and allocators run due diligence on a strategy or a manager's claim. Traders test futures, options-gamma and positioning ideas fast, data scientists run walk-forward machine learning on clean panels, risk and compliance teams audit how a model was built and tested, and educators replicate published papers on real data.

Three illustrative workflows: a futures trend rule that survives 2022, tested across 139 futures markets with published roll calendars and costs on; dealer gamma (GEX) as a risk overlay for S&P 500 exposure, where the strategy must beat its own version without the gamma signal; and replicating a published factor on 16,000+ stocks including delisted names, with fundamentals as filed and the result deflated for every variant tried. Each ends in a published strategy, a sizing rule with its own forward record, or an honest “no” with the reason.

A strategy is published only after six gates: a deflated Sharpe of at least 0.95 at the true trial count, beating its benchmark on unseen data, profitability at double costs, surviving ±20% on every setting, at least 100 trades or five years of evidence, and correlation below 0.8 to anything already published. Then the hidden hold-out decides. Failures are published too, and nothing is quietly re-run.

Typical trading apps, backtesting platforms and do-it-yourself research stacks leave most of the honesty work to the user. Alpha Factory builds it in: point-in-time data with delisted names, macro as first published, every trial counted with a deflated Sharpe, a hold-out that opens once, a public forward record that cannot be edited, a keyless start for any AI model, and data you keep.

Alpha Factory tests on the full FinzData catalogue: 16,000+ US stocks since 2000 including delisted names; 5,100+ ETFs and 123 indexes; 139 futures markets with every contract since 2007; end-of-day options chains since 2010 for 5,800+ underlyings with point-in-time dealer gamma; 22,700+ macro series with every vintage kept; COT positioning since 1986; and SEC fundamentals as filed. Bars run from tick to daily.

Alpha Factory works with every major model family, including Claude, ChatGPT, Grok, Gemini, Perplexity, Qwen, Kimi, GLM, Llama and local models, through MCP, OpenAPI, packages and plugins. It starts keyless, answers in small structured responses with the next step, gives every result a snapshot id an AI can cite and anyone can replay, and tests each model family nightly.

No result may use information that was not available on the day. Every trial is counted. Forward records are hash-chained and append-only, never edited, not even by FinzData. Every published figure is computed by the backtest engine, never written by a model. FinzData never trades on users' strategies or sells placement in a ranking, and user code runs sandboxed.

Any AI model can start with no key, no form and no card on the last six months of data, with 20 trials a day. A free email account opens the last 12 months. Buyers get unlimited trials on the full history they own, and Institutional gets everything, autopilot included. When a test needs more history, the answer names the exact product, its price and the buy link.

The long-range goal, not a forecast: Alpha Factory launches in 2026 with every trial counted; by 2031 FinzData aims to be the default data layer AI models reach for; by 2036 the referee that checks AI capital against pre-registered records; by 2041 the record of record for allocators and regulators; and by 2046 the economic memory for machines.

Try it in your AI now: “Use FinzData Alpha Factory to find a /ES strategy that survives 2022, and tell me if it is real.” Every data result carries a snapshot id (this deck: v2026-11-09), so your AI can replay any number with as_of set to the snapshot date.