Quantitative strategy development focuses on the intellectually interesting parts: signal identification, factor construction, portfolio optimisation, backtesting methodology. The infrastructure that makes a strategy run reliably in a live environment gets much less attention — until it fails at the worst possible moment.
Data pipeline reliability is the real problem
Most quant infrastructure failures aren't model failures. They're data failures. A price feed that's delayed by three minutes skews position calculations. A corporate action that isn't correctly applied to historical data corrupts backtests. A data vendor API change that isn't caught breaks ingestion silently — the pipeline runs, it just writes stale data. By the time the problem is detected, positions have been taken on incorrect information.
The solution is comprehensive data validation at ingestion, not just schema validation but value-level checks: prices within expected ranges for the instrument, volumes consistent with market conditions, timestamps correctly sequenced, no gaps in time series that would indicate a silent feed failure. Anomaly detection on incoming data catches feed problems before they propagate to downstream calculations.
“Most quant infrastructure failures aren't model failures.”
Latency and execution reality
Backtesting environments routinely understate execution costs. Slippage at the order sizes a live strategy actually trades is larger than backtested slippage. Market impact is real for strategies with meaningful AUM. Transaction costs compound over time in ways that erode strategy returns faster than backtests suggest.
Modelling execution realistically requires transaction cost analysis against actual fills. This isn't available at the design stage — it accumulates as the strategy trades. Infrastructure needs to capture fill data in enough detail to feed back into strategy parameter calibration. Strategies that work at small scale need re-evaluation as AUM grows.
Operational resilience requirements
A strategy that runs Monday to Friday needs infrastructure that survives weekend maintenance windows, vendor API changes, and market microstructure events. This means automated health checks across all dependencies, clear alerting with appropriate escalation, tested failover procedures, and — critically — a defined procedure for what the system does when it's uncertain: halt and alert rather than continue with degraded data.
The most expensive infrastructure failures we've seen in quant operations weren't technical failures. They were systems that continued operating in a degraded state without alerting, making trades based on stale or incorrect data. Designing explicit failure modes into the system is more valuable than optimising for uptime.