The Empty Repository: What a Stat-Arb PM Carries to the Next Seat in 2026

Nobody moves a stat-arb book. In September 2026 one of the largest multi-manager platforms showed what moves instead. Citadel's own site lists a global systematic equities unit, founded in 2026 and based in Irvine, California, led by a researcher who spent nearly a decade at TGS, the Irvine quant firm that eFinancialCareers describes as "super-secretive". The unit sits where he already worked. The page lists what the house supplies, "execution, optimization, and monitoring tooling", and on the hire's side of the ledger it lists a career and nothing else. A career is the whole of what a statistical arbitrage portfolio manager is allowed to carry, and a platform built a unit around one.

The same month handed every mid-frequency equity manager a number to quote. Goldman Sachs's prime desk, as Reuters reported it on 2 October 2026, had systematic equity long-short funds up 3.46 percent in September, their best month of the year, while fundamental long-short lost 0.55 percent. HFR's investable indices agreed on the direction if not the size: equity market neutral was the one equity sleeve that rose, up 0.20 percent on mean-reverting, factor-based strategies, while multi-strategy relative value lost 1.17 percent as rates jumped after the Federal Reserve's increase. If you run a market-neutral equity book at a platform or a prop firm, September was yours and the calls have started. Underneath, every one of them asks how much of that number leaves the building with you.

A portable track record in statistical arbitrage is the part of a book's risk-adjusted return that its portfolio manager can reproduce at a new firm starting from an empty repository: no code, no data history, no execution stack and no financing terms carried over, inside the time the new firm will fund. The rest of the record belongs to the building it was earned in.

Is a mid frequency equity PM move worth making in 2026?

The demand side says yes. BNP Paribas's capital introduction team surveyed 175 allocators with about $1.2 trillion in hedge funds and found 25 percent planning to add to equity market neutral and 24 percent to quant equity in the second half of 2026. Goldman's prime services team counted payouts at the top-paying multi-manager platforms at 24.5 percent of profits in 2025, against 22 percent in 2022, and described the managers at the top of them as producing Sharpe ratios above two over five years. Platforms that took in capital faster than they could seat people are short of that profile, and a market-neutral equity book with a multi-year record is a seat they will pay to fill.

The evidence on what happens after a move is less friendly. Berk, van Binsbergen and Liu found that at least 30 percent of the value mutual fund managers add comes from the firm's decisions about which of its own managers gets capital, and found no similar effect when a firm hires a manager from outside. The old firm knows something about its people that the new one has to learn. Huckman and Pisano followed cardiac surgeons who operate at several hospitals and found that a surgeon's results improve with his recent volume at that hospital and do not improve with his volume at the others. Against that, Ewens and Rhodes-Kropf tracked venture partners as they changed firms and put the partner's own human capital at two to five times the weight of the firm's. The more a job is judgement, the more of it travels. The more it is a machine with a person at the controls, the less. A stat-arb seat sits closer to the operating theatre than most of the people in it like to admit.

So the decision is narrower than stay or go. You are deciding whether to rebuild. The move is worth making when the edge is a research process you can restate without a single file, when the destination's machine is at least as good as the one you are leaving, and when the contract pays for the months in which there is no book at all.

What is the portable track record stat arb panels will pay for?

Take the Sharpe ratio apart by who owns each piece.

The signals and the code are the firm's. Trade-secret law draws that line, and a Manhattan courtroom is arguing over where it runs.

The cost function is the firm's too, and in mid-frequency equities it is most of the game. Frazzini, Israel and Moskowitz measured $1.7 trillion of live executions at one large manager and found actual trading costs an order of magnitude smaller than earlier studies had suggested. A cost belongs to whoever does the trading far more than to the signal. Novy-Marx and Velikov found that few strategies turning over more than 50 percent a month survive their own costs, which is where every mid-frequency book lives. Muravyev, Pearson and Pollet took 162 published long-short strategies and watched an average return of 0.15 percent a month become minus 0.02 percent once stock borrow fees were charged. Your net Sharpe was earned on one firm's execution, one firm's borrow and one firm's financing. Run the same forecasts through another firm and you have a different book.

The netting is the firm's as well. DeMiguel, Martin-Utrera, Nogales and Uppal showed that trading costs raise the number of stock characteristics worth trading from six to fifteen, because the trades needed to rebalance one signal cancel against another's. A signal that loses money alone can pay inside a larger book, and a platform's central book does that sum across every pod it houses. Aurum's count for the twelve months to June 2026 shows the result from outside: 22 stand-alone statistical arbitrage funds with $17.1 billion between them and a five-year Sharpe of 0.92, against 2.31 for the quant multi-strategy funds that run stat-arb beside other systematic books. Diversification explains much of that gap. Shared execution, financing and netting are in it too, and none of them is on your CV.

What is left is yours: which ideas you chose to test, how you sized them, which models you retired and when, and what you did in the weeks the book was wrong. A panel cannot audit that from your old numbers, for the reasons the piece on verified records sets out. With a stat-arb hire it can do something better: wait, and not for long.

Bailey and López de Prado's minimum track record length puts a number on the wait: the live record needed before an observed Sharpe ratio can be told apart from a lower one at 95 percent confidence. On daily returns and their simplest assumptions, an observed Sharpe of 4 needs 0.70 years to be distinguishable from 2. An observed Sharpe of 2 needs 2.73 years to be distinguishable from 1. An observed Sharpe of 1 needs 10.85 years to be distinguishable from 0.5. Each halving of the Sharpe roughly quadruples the wait, and anyone can rerun the figures from the formula in the paper.

Read the first and last of them together. A book running at a Sharpe of 4 proves it is better than 2 in about eight months of daily returns, while a discretionary manager running at 1 needs most of eleven years to prove he is better than 0.5. A stat-arb record is the least portable on paper of any seat on a platform and the quickest to earn again. That is the hiring firm's trade: it pays for a rebuild whose result it can know inside a year. The arithmetic flatters, though. The same paper shows the requirement for a Sharpe of 2 against 1 on monthly data rising from 3.24 years to 4.99 once returns carry the skew and fat tails of the hedge fund index, and mid-frequency equity has those tails. On 19 August 2026 systematic long-short funds were down about 1.4 percent by early afternoon, their worst day in more than two years, on Goldman figures reported by the Financial Times.

How much of the book is still alive after the sit-out?

Alex Morris of Selby Jennings wrote in December 2025 that twelve-month sit-outs had become common across the buy side, with some clauses running to 24 or even 36 months. G-Research put three-year non-competes on key quant staff in late 2024. In August 2026 Citadel moved higher-paid investment staff to as long as two years. In New York a bill banning non-competes for most workers and capping them at one year for anyone earning above $500,000 cleared the Senate on 3 June 2026, and the session ended that month with it still in the Assembly's Labor Committee. California is the exception, and the Irvine unit sits inside it: the state's code voids any contract that restrains a person from a lawful profession, which leaves trade-secret law as the only fence there.

Then the cost that is particular to this seat. A discretionary manager's views were going to change during the sit-out anyway. Yours have a measured life. Di Mascio, Lines and Naik followed 1.15 million institutional trades and found alpha of 37 basis points in the first month after a purchase, declining to effectively zero by month twelve. Bowles, Reed, Ringgenberg and Thornock showed in the Journal of Finance in 2024 that the returns to published signals are concentrated in the first month after the information is released and decay soon after. A sit-out of one to two years outlasts every position you held and most of the forecasts behind them. What survives is the way you find the next ones, and that was the portable part to begin with. It travels, too. Jacobs and Müller tested 241 anomalies in 39 stock markets and found the United States the only country with a reliable decline in returns after publication. A research process built in the most arbitraged equity market in the world gives up less when it is pointed at a market with fewer people running it.

Then the rebuild, where the public record has no number at all. No platform publishes how long a systematic team takes to go live, and anyone who quotes you a figure is quoting an anecdote. What can be fixed is in the contract. The guarantee has to cover the sit-out and the build. eFinancialCareers reported guarantees of $30 million and $50 million paid to hires in 2025, and the money must usually be repaid if you join a competitor inside two years, and is yours to keep if you are stopped out, so the lock runs through the same period in which the new record does not yet exist. The payout percentage means little until you know which costs come off before it. On a February 2025 count about 80 percent of multi-strategy firms pass expenses through to their investors, on lists running from compensation to AI tools, so the contract should say which of your data, compute and researchers are charged to your book before the payout is calculated. The ramp should be a written schedule, with capital stepping up at stated live Sharpe milestones and the stop measured against the capital allocated at each step. Bailey and López de Prado's arithmetic tells you which milestones are statistically honest.

Can you take the team if you cannot take the code?

The research says take the team. Groysberg and Lee split star analysts who changed firms into those hired to reinforce something the new firm already did and those hired to start something it did not. The second group suffered a decline in performance that lasted, and kept more of what they had when they moved with a group of colleagues. A stat-arb manager hired to found a unit is the second kind of hire by definition. The team is the nearest thing he has to a portable machine: the researcher who remembers why the third version of a model was abandoned, the developer who has built the simulator once already.

The law says take nothing else, and the Southern District of New York is deciding where "nothing" ends. The criminal case over source code allegedly taken from Headlands Technologies, which survived a motion to dismiss in June 2026, was adjourned on 9 September 2026 to a trial date of 2 November 2026. Two orders from the same month matter to anyone planning a rebuild. On 22 September 2026 the court let the Government test the defence's position that the code is "not novel in its logic, functionality, or trading insights because it exists at these other firms". The next day it set briefing on "the definition of a trade secret as relevant to combinations or compilations", which asks whether an arrangement of individually ordinary parts can be a secret as a whole. That is the stat-arb rebuilder's own question. Most components of a mid-frequency book are in a textbook somewhere; the arrangement is the book. The defendant is presumed innocent and the case says nothing about any other hire. Its answers will still describe the fence every team rebuilding from memory works inside, in California as much as in New York.

When does the new record exist?

Run the calendar forward from the 2026 pay round. For most of the industry, bonuses for 2026 are paid in the first quarter of 2027. Resign after that payment with a twelve-month sit-out and you sit down at the new firm in the first quarter of 2028; with twenty-four months it is 2029. Assume the rebuild takes no time at all, which it will not. A book that comes back at a Sharpe of 4 still needs until late 2028 to prove it is better than 2, and one that comes back at 2 needs until late 2030 to prove it is better than 1. The offer you sign before that bonus is priced on a record that cannot exist before then, and a two-year lock on a guarantee that starts in early 2028 runs out in early 2030, in the middle of that range. Ask for the ramp milestones to be dated inside the lock.

Two things on the public record move before any of that. The Headlands trial is set to open on 2 November 2026, and its rulings on compilations should say how much of an arrangement a rebuilder may reproduce. And capital for 2027 is being argued over in the same weeks, by platforms that watched a rival decide in September that a systematic equity seat was worth a new unit in the hire's own town. If you are going to make the move, the first file in the empty repository is the plan for filling it: what you will test, in what order and on what data, written after you have left and from nothing but your own head. Until the end of 2028 at the earliest, it is the only track record you will have.