Writing a Research Note

Most research notes are written backwards. A researcher spends three weeks on a project, and the note follows the order the work happened: data, then method, then results, then a conclusion buried in the last paragraph. That order makes sense to the person who lived it and to almost nobody else. A reader — a PM deciding whether to fund the next stage, another researcher deciding whether to build on it — needs the conclusion first, so they can decide in thirty seconds whether the rest is worth their time at all.

The order that actually works

The answer, in one sentence, first. "Twelve-one month momentum earns a 3.1% annualized decile spread after costs in the top 1,500 US names, 2010–2024, uncorrelated with anything currently in the book." Not "we investigated whether momentum..." — the reader should know the outcome before the first full paragraph ends.

The decision it implies, stated explicitly. "Recommend building this out as a candidate signal at up to $150m capacity" or "recommend killing — capacity too small to matter." A note that reports a result without stating what it implies forces every reader to do that translation themselves, inconsistently.

The scope, immediately after. Universe, period, cost model, in one or two lines — exactly the four things a well-framed question specifies. This lets a skeptical reader immediately spot the boundary of the claim: "does this hold outside the top 1,500? Unknown, not tested," rather than discovering the boundary three paragraphs in.

The evidence, only now. The tables, the robustness checks, the caveats. This section can be as long as it needs to be, because by this point the reader has already decided whether they need the detail or just the headline.

What would change the conclusion. The single most useful closing line in any note: what result, if it showed up next, would reverse this recommendation. It tells the next researcher exactly what to check before building on it, and it's the fastest way to catch a note that's actually overconfident.

What to leave out

A note is not a diary of the research process. The three approaches that didn't work, the data cleaning war stories, the dead ends — these belong in a private log or an appendix, not the main note, unless one of them changes how much to trust the headline result. A note cluttered with process detail trains readers to skim, and a skimmed note is read exactly as carelessly as a bad one.

A short example, start to finish

"Recommendation: build out. Short-interest reversal earns a 2.8% annualized decile spread after realistic costs, Russell 3000, 2012–2024, correlation 0.15 with the existing quality composite. Estimated capacity ~$180m before the edge halves.

Scope: Russell 3000, monthly rebalance, costs from the desk's standard model, held-out final two years not yet checked.

What would kill this: if the effect concentrates in the bottom size decile once broken out (untested), or if held-out period performance falls materially below 2%."

Three sentences, and a reader who trusts the desk's standard cost model and benchmarking harness can make a real decision from them alone — request the full note only if they need to check the scope boundary or the robustness detail themselves.