The Research Funnel: From Idea to Live Capital

A research desk that logs a hundred ideas in a year will put one or two of them in front of real capital. That ratio is not failure; it is the job. The interesting question is not how to raise the hit rate — you mostly can't — but how to arrange the work so the ninety-eight failures cost days instead of months. That arrangement is the funnel.

The funnel is a cost-ordering, not a process document

Every gate exists for one reason: it kills a certain fraction of ideas for a certain price. A good funnel orders the gates by price. The half-hour question that eliminates 60% of candidates goes first; the three-week capacity study that eliminates 20% goes last. Reverse the order and your annual output drops by an order of magnitude, on exactly the same ideas.

The gates

Stage	The question it answers	Time per idea	Survives
Triage	Is there a mechanism? Do we have the data? Who is on the other side?	30 minutes	~40%
Quick-and-dirty pass	Does the effect exist at all, in the sign we predicted, with no cleaning?	half a day	~35%
Signal build	Does it hold with point-in-time data, proper lags, neutralisation, and a sane decay profile?	1–2 weeks	~40%
Portfolio test	Does it survive turnover, borrow, spread and impact at the size we would trade?	1–2 weeks	~50%
Robustness + holdout	Is it there across subperiods, universes, and on the sealed sample — one shot?	3–4 days	~65%
Paper / shadow	Does the live plumbing reproduce the backtest on fresh data?	1–3 months	~70%
Live, small	Does it work when it is our own flow moving the price?	3–12 months	~50%

Two things matter more than the exact numbers. Over half of everything dies in gates that together cost less than a day. And survival does not rise steadily down the funnel — the live gate kills roughly half of what reaches it, because no backtest contains your own market impact, your own fills, or the three other funds that started trading the same thing that quarter.

One idea through the pipe

Triage. Someone reads that clusters of open-market insider buying — three or more officers buying the same stock in a fortnight — predict returns. Mechanism: insiders know more, and buying with their own money is costly signalling, so this is an informational edge. Data: Form 4 filings with filing timestamps, which we have. Other side: whoever sold, plausibly uninformed. Passes, ten minutes.

Quick-and-dirty. No neutralisation, no cleaning: for every cluster event since 2012, compute the 20-day forward return minus the sector median. Average is +1.4%, on 3,100 events, sign as predicted. Half a day. Passes.

Signal build. Now do it properly. Use the filing timestamp, not the transaction date — insiders have two business days to file, and using the trade date is a look-ahead of one to two days that turns out to be worth a third of the effect. Drop 10b5-1 pre-scheduled purchases, which carry no information. Neutralise against size and sector. The effect drops to +0.7% over 20 days and decays to nothing by day 35. Still passes, but it has halved. This is normal.

Portfolio test. Cluster events are rare: about 260 a year across the universe, so the book holds maybe 40 names at a time, and turnover is high because the horizon is a month. Net of 25 bps round-trip and impact at our size, the Sharpe Ratio lands at 0.55 with a capacity ceiling around $120m. Marginal — it goes forward only as a blend component, not a standalone book.

Robustness. Sector-by-sector the effect is positive in nine of eleven sectors; period-by-period it is much weaker after 2019, which is when the data vendor started selling a pre-packaged version of it. That last observation is worth more than the backtest: it tells you the decay is crowding, not noise (Factor Crowding).

Verdict: small weight in the blend, revisit in a year. One idea, four weeks, honest answer "a bit, and fading" — which is what most surviving ideas look like.

Three ways the funnel breaks

Gates that never kill anything. If a stage has a 95% pass rate, it is a ritual, not a gate. Either sharpen its criterion or delete it and save the time.

Skipping to the expensive stage. The most common failure on junior desks: a full backtest infrastructure spun up for an idea that a thirty-minute scatter plot would have killed. Enthusiasm is not a reason to reorder the gates.

Recycling corpses. An idea killed at stage three comes back three months later with one parameter changed. That is not a new idea; it is the same test again, and it belongs in your count of variants tried (Hypothesis-First vs Data-First Research). Keep the kill log and check it before starting.