The Funnel
Seven stages of elimination
Every strategy that runs capital passed through the same gauntlet. Most are destroyed at the first gate. The ones that survive the seventh are the ones we trust.
828,000+
Configurations tested173 archetypes across 4,784 equities, plus scanner permutations
26,000
Hypothesis tests generatedScanner output: signal candidates with measurable edge
1,704
Showed persistent signalPassed initial screening with statistical significance
158
Formally backtestedFull walk-forward methodology on 16 years of data
66
Passed holdout testZero data overlap with discovery period
26
Verified independentUncorrelated with existing engines
11
DeployedRunning capital in the portfolio
Elimination Rate
99.99%
of everything we test gets eliminated. The 0.01% that survives runs your capital.
The Pipeline
Seven stages, one standard
Each stage is designed to kill. A strategy that passes Stage 3 has already survived more scrutiny than most hedge funds apply to their entire portfolio. A strategy that passes Stage 7 has been tested on data it has never seen, measured against every other engine for independence, and observed in live-market conditions before receiving a single dollar of allocation.
Hypothesis Generation
Factory archetypes, scanner signals, academic literature, and structural market analysis produce raw candidates. The autonomous research machine generates thousands of hypotheses per cycle. No human bias in the search -- the system tests everything.
Initial Backtest
Does the signal exist? A rapid screen on historical data to determine whether the hypothesized edge produces any measurable return. Most candidates die here -- no signal, wrong direction, or signal too weak to overcome transaction costs.
Walk-Forward Validation
Does it work on unseen data? The backtest is repeated with strict temporal separation -- the model is trained on one window and tested on the next, rolling forward through the entire data history. This eliminates strategies that only worked because they saw the future.
Holdout Test
Zero overlap with discovery data. A final, completely independent test on data that was never used during any part of the research process. Not the same data in a different window -- entirely separate data that the strategy has never seen. This is the test that caught 68 false positives in a single strategy before capital was deployed.
Correlation Check
Is it independent from existing engines? A strategy with Sharpe 2.0 is worthless if it moves in lockstep with three engines already in the portfolio. Every candidate is measured for pairwise correlation against all deployed engines. Only strategies with genuinely independent return streams advance.
Portfolio Impact Simulation
Does it improve the portfolio? The candidate is added to the existing portfolio in simulation. If the portfolio Sharpe ratio does not increase, or if drawdown characteristics worsen, the strategy is rejected -- even if its standalone numbers are strong. The portfolio is the unit of measurement, not the strategy.
Paper Trading Deployment
Live market observation with zero capital at risk. The strategy runs against real-time data, generating real signals and recording real fills at market prices. Slippage, latency, and execution quality are measured. Only after a sustained period of live observation that matches backtest expectations does the strategy receive allocation.
"A strategy that survives seven stages of testing designed to kill it is not lucky. It is structural. Luck does not replicate across sixteen years of market data, multiple asset classes, and live observation."
The Factory
An autonomous research machine
Most quantitative research is done by humans sitting at desks, running backtests during business hours, and reviewing results over coffee. We built a machine that never sleeps.
Autonomous Research Pipeline
Every night at 2:00 AM, the factory wakes up. It scans the entire instrument universe -- equities, crypto, futures, commodities -- across 173 archetype configurations. It generates hypothesis tests, runs backtests, applies walk-forward validation, and ranks the survivors. By morning, results are waiting. No human in the loop during scanning. No bias in the search. No fatigue, no shortcuts, no pet theories.
173
Archetype configurations
4 classes
Equities, crypto, futures, commodities
2:00 AM
Nightly pipeline execution
0
Humans in the scanning loop
The factory does not replace human judgment. It replaces human limitations. A team of researchers can test perhaps five to ten hypotheses per week. The factory tests thousands per cycle. The human role shifts from generating and testing ideas to evaluating the survivors -- the highest-leverage use of human intelligence in the research process.
Infrastructure
The data underneath
Scale is not a number on a slide. It is the infrastructure that makes the number real. Behind the funnel sits a data lake, a signal library, and a production codebase built to handle the volume.
93M
Signal rows in the data lake
3,792
Instruments monitored
30K+
Lines of production code
16yr
Historical data depth
140+
Unique measured signals
Every signal row represents a measurement -- a data point that was computed, stored, and made available for the factory to test. 93 million rows means 93 million observations of market behavior, each one a potential building block of a strategy. The breadth of instruments ensures that the system is not over-fitted to a single market. The depth of history ensures that strategies are tested across multiple market regimes -- bull, bear, crisis, recovery, and everything in between.
"The research funnel is not a process we invented. It is what honest research looks like when you refuse to deploy anything that cannot survive systematic destruction."
828,000 configurations enter the top. 11 engines emerge at the bottom. The distance between those two numbers is the distance between hope and evidence. Every strategy that runs your capital crossed that distance.