Working With a Portfolio Manager

The most useful thing to understand about a portfolio manager is that they are not asking your question. You ask "is this signal real?" They ask "should I take risk off something I already own and put it here, given that I answer for the result in November?" Those have different answers often enough that a researcher who misses the distinction spends years confused about why good work gets turned down.

Two seats, two jobs

The researcher owns evidence: the number is honest, the test was fair, the caveats are stated. The PM owns risk and capital: a book that survives a bad quarter, stays inside its mandate, and can be explained to whoever allocated the money.

The practical consequence is that the PM is allowed to reject a true result. A signal can be genuinely predictive and still be a bad addition — it correlates with what the book already runs, its drawdowns land with everything else's, it needs borrow the desk cannot get, or it caps out too small to justify the operational risk. None of that is a criticism of your research, and hearing it as one is the standard early-career mistake here.

What a PM is actually asking

Four questions come up in almost every meeting. Have numbers for all four before you walk in.

"What does this add to what I already have?" Not the standalone Sharpe — the marginal contribution. Give the correlation of the new signal's daily P&L to the book, and the book's Sharpe with and without it. A standalone 1.8 running at 0.75 correlation with the momentum sleeve may add nothing; an uncorrelated 0.6 can be worth more.

"When does it stop working?" Name the conditions: "weak when cross-sectional dispersion is in the bottom quartile," or "needs at least four analysts covering the name." A researcher who cannot describe the failure regime has not understood their own signal, and the PM knows it.

"How big can it go?" Capacity in dollars, assumptions stated: at what participation rate, at what impact, and what the Sharpe becomes at half and double that size (Portfolio Capacity).

"What does the worst quarter look like?" Give the backtest's worst drawdown, and then the honest caveat: a 12% worst drawdown over eight years has not proven 12% is the limit, only that nothing worse happened in eight years.

What you say and what gets heard

You say	The PM hears	Say this instead
"The Sharpe is 1.8"	"Selected from many variants"	"1.8 in-sample, 1.1 on the sealed holdout; I tested 14 variants"
"It works across all sectors"	"Sectors were the only cut he tried"	"Positive in 9 of 11 sectors; the negatives are utilities and REITs, both rate-driven"
"Costs are handled"	"Costs are a plug number"	"25 bps round-trip, from our own fills on similar names; at double that the Sharpe goes to 0.6"
"It's uncorrelated with the book"	"Over the full sample, maybe"	"0.05 full-sample, but 0.4 in the worst decile of market days"
"I need more time"	"It's not working"	"Two more weeks on the capacity study; the alpha result is done and it's 0.7"

PMs discount confident statements automatically, because they have been shown a lot of backtests. The way through is not more confidence but volunteering the weakness first. A researcher who leads with "here is the part I am least sure about" gets believed on everything else.

A twenty-minute meeting

You have a signal from earnings-call language, Sharpe 1.8 in-sample. The version that works:

Minute 1. The claim, in one sentence, with the caveat attached: "Language changes in earnings calls predict a two-week drift; I get a net Sharpe of 1.1 on the holdout, capacity around $200m, and it is correlated 0.3 with our post-earnings drift sleeve."

Minutes 2–8. The evidence in the order a sceptic wants it: decay curve, subperiod stability, the cost assumption and where it came from, the holdout result. One chart each.

Minutes 9–14. The caveats, unprompted. "Two things bother me. It is a third weaker after 2021 — crowding or noise, I cannot tell with this sample. And a fifth of the P&L comes from names under $2bn, where our fills are worse than the model assumes."

Minutes 15–20. The ask, specific and small: "I would like $25m for six months, sized at half the model weight, and I will bring back live-versus-backtest attribution monthly."

Note what is missing: no build-up, no defence of a methodology nobody attacked, no request for a large allocation. The small ask converts an argument about whether you are right into an experiment that settles it, and PMs say yes to experiments far more readily than to conclusions.

Credibility is the actual asset

The researchers who get capital are not the ones with the best backtests but the ones whose numbers hold up. Three unglamorous habits build that: state the count of variants you tried without being asked; show live-versus-backtest every month, including the bad ones; and say "I don't know" precisely — "I cannot separate crowding from noise with eight years of data" — rather than as an apology.

The reverse matters too. When a PM asks for something you think is wrong — "just drop 2008, it distorts everything" — the answer is neither refusal nor compliance. Run it both ways, show both, say which you would use and why. You keep the evidence seat, they keep the decision seat, and re-litigating that split is how the relationship stops working.