Demand forecasting for infrastructure: lessons from predicting traffic spikes

We manage infrastructure scaling for 15 e-commerce clients. Demand forecasting for infrastructure is fundamentally different from demand forecasting for inventory:\n\n**Lead time is minutes, not weeks.** A traffic spike happens now. You need to have scaled 5 minutes ago. This means predictive scaling, not reactive scaling.\n\n**The cost of overshoot is low. The cost of undershoot is catastrophic.** An extra 20 servers for 2 hours costs 0. A 10-minute outage during Black Friday costs 0,000.\n\n**External signals matter more than historical patterns.** A press mention, a viral tweet, or a competitor going down can drive 10x traffic with zero historical precedent.\n\nWhat we've built: a multi-signal forecaster that watches traffic trends, social mentions, competitor status, and scheduled events. It pre-scales 15 minutes before predicted spikes and scales down conservatively after the event.\n\nResult: 99.97% uptime across all clients during the last holiday season.