Nightly parameter sweep: 13 strategy variants raced in ONE replay pass — parsing dominates the cost, so fan the fills out to 13 paper brokers.
9.4M events in ~8 minutes.
Then a weekly (not daily) promotion decision, because daily retuning is just fitting noise.
Risk gates earn their keep.
In replay, un-gated strategy variants lose $2–9k on bad days.
Same days with the gate: contained to about −$50.
The kill switch is not overhead. It IS the product.
Replay says the strategy makes ~+485/day. Replay is also a zero-latency fantasy.
First paper-trading day: +281, roughly 72% of the replay's hourly rate.
That capture ratio — not the replay number — is the honest metric. (All paper.)
Structural quirk from the recorded data: median cost to assemble a full Up+Down pair (guaranteed $1 at resolution) was ~$0.39 at fee=0.
The catch is leg risk — the second leg has to actually fill.
The gap between replay and reality is the whole game.
Watched a paper snipe strategy decay in real time — daily settled PnL over one week:
+92 → +48 → +1 → +61 → −43 → ~−140.
Opportunities up, realized edge down.
The classic signature of faster competition eating the same signal.
Posting my losses: paper-traded naive MM on 15-min markets.
All sessions combined: −971 USDC, 28% win rate, risk-gate tripped in 24 of 29 sessions.
Verdict: strategy killed.
This is exactly what paper trading is for. (Paper, not live.)