Splitting Time Between Maintenance and New Alpha
A research team's time is a fixed budget split between maintaining and monitoring existing signals and hunting for new ones, and underinvesting in the first is a slow, silent way to lose returns that already exist.
Every research team faces the same allocation question with no formula answer: how much time goes into finding brand-new signals versus maintaining, monitoring, and quietly improving the ones already live. New alpha is the visible, exciting work that gets credit in performance reviews and conference talks; maintenance — checking that a live signal's data feed hasn't drifted, that its edge hasn't decayed, that a small parameter refresh could meaningfully help — is invisible when done well and looks like nothing happened.
The risk of underinvesting in maintenance is not dramatic; it's a slow leak. A live signal's edge erodes gradually as the market adapts or as its underlying data source degrades, and if nobody is specifically assigned to watch for that, the strategy keeps running exactly as before while its actual profitability quietly declines — the team only notices once the decline is large enough to show up clearly in aggregate P&L, by which point months of avoidable underperformance have already happened. Because new-alpha research and maintenance compete for the same researcher-hours, teams that don't explicitly budget time for maintenance tend to default to zero, simply because new work is more rewarding to prioritize in the moment.
The common practical fix is treating maintenance as its own line item with dedicated time (a fixed percentage of the week, or a rotating on-call researcher) rather than leaving it to whatever time is left over after new-idea work, which in practice is usually none.
Because maintenance work is invisible when done well and new-alpha work is what gets recognized, teams tend to underinvest in maintenance by default unless it is given its own explicit time budget — and the cost of skipping it shows up as a slow, hard-to-notice decline in existing signals' live performance rather than any single visible failure.
Further reading
- Chan, Quantitative Trading, ch. 8