What you'll take away
- Reproducibility is becoming table stakes for compliance, and reproducibility is impossible without determinism.
- Many optimizers are quietly non-deterministic — parallelism, timeouts, and floating-point order make the same inputs return different trades.
- Determinism is a feature you design in, not a property you hope for. Done right, it turns the compliance team from a blocker into an ally.
Start with a question that sounds easy and usually is not: if you re-ran last Tuesday's rebalance on last Tuesday's data, would you get exactly the same trades? For a surprising number of production systems, the honest answer is "probably, mostly, unless something was under load." That hedge is the whole problem. An investment decision you cannot reproduce is an investment decision you cannot fully explain, and one you cannot explain is one you cannot defend to a client, an auditor, or a regulator who asks the most reasonable question in the world: why did this account trade this way on this day?
Why optimizers drift
Non-determinism creeps into numerical systems through several doors, and most of them are doors you opened for good reasons. Parallelism is the biggest: when work is split across many cores or a GPU, the order in which partial results combine can vary run to run, and floating-point arithmetic is not associative, so a different summation order can produce a slightly different number. Usually the difference is tiny — but "tiny" is enough to flip a marginal decision, and a flipped decision means different trades. Timeouts are another door: a solver that returns "the best I had when the clock ran out" returns something that depends on machine load, which is the definition of non-reproducible. Random initialization, unstable tie-breaking, and dependency-version drift each add their own quiet variance.
Figure 1 — Same inputs, different outputs: the non-determinism trap
If runs A and B can disagree, you cannot promise an auditor that any given decision is re-derivable. The fix is to remove the sources of variance by design.
The regulatory direction of travel
The supervisory expectations around models have been tightening for years. In banking, the model-risk-management guidance commonly referred to as SR 11-7 set an early, influential standard: models should be documented, validated, and governed, with their behavior understood and reproducible. The spirit of that guidance has steadily diffused outward into asset and wealth management, where the questions are now familiar in any due-diligence process — how is the model validated, can decisions be reproduced, who governs changes, what is the audit trail. A firm that can answer "yes, deterministically, here is the re-derivation" has a structural advantage in those conversations. A firm that has to explain why its optimizer is "basically reproducible" has a problem it has not finished solving.
Determinism as a design choice
The good news is that determinism is achievable; the demanding news is that it has to be engineered deliberately. It means controlling the order of parallel reductions so floating-point results are stable, removing wall-clock timeouts from the decision path so load cannot change the answer, fixing seeds and tie-breaking rules, pinning dependency versions, and — the capstone — content-hashing inputs and outputs so that any past decision can be located and re-run bit-for-bit. When a system is built this way, "re-derive the trade from March 14th" stops being a fire drill and becomes a lookup.
The strategic payoff is that this inverts the usual relationship with compliance. Most optimizers worry compliance teams precisely because they are non-reproducible — the team cannot independently verify what happened, so they treat the system as a risk. A deterministic, content-hashed, re-derivable system gives compliance exactly what it wants: the ability to reconstruct any decision on any date and confirm it followed the stated rules. The blocker becomes an ally, and the audit that everyone dreads becomes a query. That is not just good engineering hygiene; in a market where trust is the scarce resource, it is a selling point.
Reproducibility you can demonstrate. See how deterministic, content-hashed outputs make any decision re-derivable — and bring your compliance team to the first conversation, not the last.
Request a pilot →References & further reading
- Board of Governors of the Federal Reserve System & OCC, Supervisory Guidance on Model Risk Management (SR 11-7 / OCC 2011-12).
- D. Goldberg, "What Every Computer Scientist Should Know About Floating-Point Arithmetic," ACM Computing Surveys, 1991 — on why summation order matters.
- Asymmetry Computing, Build vs. buy: the true cost of an in-house optimization stack.