A market is the most honest reviewer a model will ever meet. It charges fees, slips your fills, drops data and answers slowly at the exact moment your signal fires. Building a signal-to-execution platform across more than fifty assets and several venues taught lessons that apply far beyond trading.
First, accuracy is the wrong question. A gradient-boosted buy signal trained cleanly and still had negative expected net return at every threshold once fees and slippage were counted — precision around 0.20 against a 0.25 break-even. The platform was ready to trade; the discipline was to decide that it was not allowed to. Rule: no execution until expectancy is positive net of costs, out of sample, no exceptions.
Second, parity must be certified, not assumed. An undeclared training-window difference can make two “identical” runs train different artefacts. Training was made deterministic, artefacts byte-tested, feature and inference parity certified with strict equality on signals and trades, and the whole thing turned into a merge-blocking CI gate — committee-reviewed, 8.9 out of 10.
Third, the safety path must be simpler than the system it protects. The kill switch is a single PostgreSQL row polled every 500 milliseconds; three failed polls and the engine behaves as engaged; automatic channels can engage but never disengage; it halts new signals but never cancels in-flight orders, because the exchange API has no place inside a safety path. Eight invariants, a chaos-test plan, and live trading gated until it is proven.
Every enterprise AI programme has a market somewhere: a regulator, an auditor, a customer who churns. The question is whether you let it review your model before or after the money is committed.
Was this useful? One click, no account.
Want to take it further? Contact us about this · Share on LinkedIn