Learn→Time Series Forecasting→Statistical Models (ARIMA, Exponential Smoothing)→4 of 6

Time Series Forecasting • Statistical Models (ARIMA, Exponential Smoothing)Medium⏱️ ~3 min

Production Architecture for Statistical Models

Key Point
Statistical models are embarrassingly parallel—each series fits independently. Millions of series can be forecast by distributing work across machines.
BATCH TRAINING PIPELINE
Flow: ingest data daily → fit one model per series → store parameters and forecasts. Fitting one model: 10-100ms. For 1M series on 100 workers: ~20 minutes total. Bottleneck is data I/O, not model fitting.
MODEL STORAGE
Store parameters, not raw data. ETS needs ~50 bytes/series. For 10M series, storage is ~500MB. Pre-compute forecasts for next N periods; store in key-value store for sub-millisecond serving.
💡 Insight: Pre-computed forecasts turn prediction into a lookup. At serving time, just fetch from cache—no inference needed. Latency: <1ms.
FRESHNESS VS COST
Pre-computed forecasts get stale as new data arrives. Daily retraining is typical. Hourly costs 24x more. Some retrain only when drift detected (accuracy drops below threshold).
AUTO MODEL SELECTION
Not all series need the same model. Per-series selection: fit ETS, ARIMA, baseline. Pick winner by cross-validation error. Unpredictable series fall back to baseline when complex models do not help.
⚠️ Trade-off: Auto-selection adds 3x compute but improves accuracy 5-15% on average. Worth it for high-value forecasts; skip for low-stakes.

💡 Key Takeaways

✓Statistical models are embarrassingly parallel—each series fits independently

✓Fitting takes 10-100ms per series; 1M series on 100 workers = 15-30 minutes

✓Store parameters (~50 bytes/series), not raw data; pre-compute forecasts for <1ms serving

✓Daily retraining balances freshness and cost; hourly costs 24x more

✓Auto-selection (ETS vs ARIMA vs baseline) improves accuracy 5-15%

📌 Interview Tips

1Explain that pre-computed forecasts turn inference into a cache lookup

2Discuss freshness vs cost tradeoff: daily retraining is typical, hourly is expensive

← Back to Statistical Models (ARIMA, Exponential Smoothing) Overview