A Stage-Gate Research Process

Give a researcher $50 million on the strength of a backtest and you are betting the whole allocation on a single, unverified number. Give them nothing until the strategy has traded live for a year and you will starve every idea before it has a chance to prove itself, since almost nothing survives to a year of live trading without some capital along the way to fund the infrastructure and data needed to test it properly. A stage-gate process solves this by breaking research into a small number of fixed stages, each with a named owner, a defined exit test, and a capital number attached — so the amount of money and infrastructure behind an idea grows only as the evidence for it does.

This is the same shape as The Research Funnel: From Idea to Live Capital, but a stage-gate process makes the funnel's checkpoints binding: a strategy cannot cross a gate on enthusiasm, only on meeting that gate's stated bar.

The five gates

Stage	What happens	Typical exit bar	Capital behind it
1. Intake	Idea is logged, screened against known strategies and current book	Idea is novel enough or improves an existing signal	None
2. Backtest	Full historical simulation against the firm's The Research Standards Document	Meets the firm's minimum out-of-sample Sharpe and cost-adjusted return bar	None — research time only
3. Paper trading	Signal runs live against real-time data, no capital, orders simulated	Live paper Sharpe within a stated band of the backtest's out-of-sample figure	None
4. Live pilot	Small real capital, tight risk limits, close monitoring	Costs, slippage and capacity match the paper estimate; no unexplained drawdown	1–5% of eventual target size
5. Full allocation	Strategy joins the regular book, sized by the standard risk process	Ongoing, not a one-time bar — subject to the same drawdown and decay monitoring as every live strategy	Full target size

Each gate is a kill point, not just a milestone. A strategy failing a gate does not get "one more look" by the researcher who built it — it stops, and restarting it requires going back to an earlier stage with a specific, named fix, decided by whoever owns that gate rather than by the researcher's own judgment.

Worked example: a volatility carry idea moving through the gates

A researcher proposes selling short-dated index variance against a delta-neutral hedge. At intake, the desk checks it isn't a near-duplicate of an existing carry book — it passes, since the existing book trades single-name variance, not index. At the backtest stage, the strategy clears the firm's cost model and out-of-sample bar, showing a Sharpe of 1.3 net of a conservative transaction-cost assumption over eight years, including the 2018 and 2020 volatility spikes.

At paper trading, run for three months against live quotes, the signal's turnover comes in noticeably higher than the backtest implied, because the backtest had used end-of-day marks that understated how often the position needed rebalancing intraday. The paper Sharpe, adjusted for the true turnover, comes in at 0.7 — below the band the gate requires. The strategy does not advance. It returns to the backtest stage with one specific fix required: re-run the simulation using intraday rebalancing costs, not end-of-day marks. Six weeks later the revised backtest, now honestly reflecting turnover, still clears the bar at a reduced but real Sharpe of 0.9, and the strategy re-enters paper trading. It passes this time, moves to a live pilot at 2% of target size, and after four clean months without an unexplained drawdown, is sized to its full allocation.

Nothing about this sequence involved a bad decision at any single point. What it demonstrates is the entire purpose of the gate structure: the flaw in the backtest's cost assumption was cheap to catch at the paper-trading stage, because only research time was at risk, rather than expensive to discover after real capital had been committed to a strategy whose live turnover nobody had actually measured.