Fewer strategies reached production, and the ones that did behaved in live trading roughly as the research had said they would.
Signal research lived in individual analysts’ notebooks. Each carried its own cost assumptions, its own universe definition and no record of how many variations had been tried before the reported result. Strategies that reviewed well consistently underperformed once funded, and the investment committee had no way to distinguish a genuine edge from the best of many attempts.
We built a shared research platform with three properties the previous process lacked: one transaction cost model applied to every backtest, an out-of-sample window held back and released only once, and automatic recording of how many configurations were evaluated so the reported statistic could be deflated accordingly. Capacity analysis was made part of the standard output rather than an afterthought.
Fewer candidates survived review, which the committee treated as the platform working rather than failing. Live performance now tracks the research estimate within a band the firm agreed in advance, and every promoted strategy carries its trial count and decommissioning rule on the record.
A median tick-to-trade that looked competitive was hiding a 99th percentile that decided the profitability of the book.
