Adapting an Academic Result Into a Tradeable Book

A junior researcher successfully replicates a published anomaly — the number holds up, close to the paper's, on the desk's own point-in-time data. The natural next step feels like: ship it. That step is usually two months too early. A replicated academic result and a strategy the book can actually run are different objects, and the gap between them is where most of the real work in this part of research happens.

What the paper never had to solve

Cost, at the desk's actual size, not the paper's. Academic transaction cost models are typically flat, small, and blind to the desk's own footprint. A signal that trades the bottom size deciles heavily needs a realistic, size-aware cost model before its true after-cost return is known — see Designing a Signal With Costs In From the Start. This is usually the single biggest source of shrinkage between a paper's headline number and a tradeable one.

Capacity, which the paper has no reason to report. An academic anomaly is measured on an equal-weighted or value-weighted portfolio with no constraint on how much capital could actually take those positions without moving the names. A signal with a strong decile spread but a capacity of $20m isn't a strategy for most desks — it's a curiosity. Estimating capacity means modelling how the signal's own trading would erode the spread as size grows. See Capacity-Constrained Backtesting.

Fit against what's already in the book. A paper is written in isolation; a desk's book is not. The question isn't "is this signal good," it's "does this signal add anything once everything currently trading is accounted for" — see Incremental Alpha of a New Signal. A great standalone signal that correlates heavily with three signals already live adds little marginal value, however clean its paper looked.

Turnover discipline the paper didn't need. Academic rebalancing is often monthly or quarterly by convention, not because that's the frequency a real trading process needs. Re-deriving the right rebalance cadence — balancing signal decay against costs — is implementation work the paper simply never had to do.

The order that avoids wasted effort

Do these roughly in order of how cheap they are to check, so a bad result kills the project before the expensive steps: re-cost first (cheap, often decisive), then check correlation against the existing book (cheap, a library lookup), then estimate capacity (moderate effort), then, only if all three survive, build out the full production-grade signal with proper turnover and rebalance logic (expensive). Doing them in the wrong order — building the full production pipeline first, discovering afterward that costs eat the whole edge — is the most common way weeks get wasted on this stage.

A worked case

A momentum-reversal paper replicates cleanly: 12-1 month momentum, decile spread intact on the desk's point-in-time data, close to the published number. Re-costing at the desk's actual size cuts the spread by roughly a third — still positive, still worth pursuing. A correlation check against the alpha library shows 0.35 correlation with an existing trend signal already live — meaningful, but not disqualifying; the paper's specific 12-1 formulation captures something the existing signal doesn't fully overlap with. Capacity estimation, modelling the signal's own market impact as size grows, puts a comfortable ceiling near $300m before the edge halves — well above what this desk would allocate to a single signal. Only at this point does it make sense to build the full production version, with a rebalance schedule tuned to the signal's actual decay rate rather than the paper's arbitrary monthly convention.