Forecasting Models

How to forecast service load: simple growth, seasonal decomposition, Prophet / ARIMA, and when ML is overkill.

Overview

Simple models usually win

Capacity planning needs forecasts, and forecasting is its own discipline. The good news: most production services don't need fancy models; simple ones work.

The forecast hierarchy, lowest to highest sophistication

1. Linear extrapolation — "Last quarter we grew 20%. Forecast 20% next quarter." — Steady, organic-growth services — Fails on saturation, viral growth or seasonality
2. Linear with seasonality — 20% growth applied to the same daily, weekly and yearly shape: peak hour 5× off-peak, Black Friday 3× a normal day — Most consumer services; standard in capacity sheets — Regime changes
3. Exponential smoothing (Holt-Winters) — Trend + seasonality + level, weighted by recency; in most spreadsheets and Python statsmodels — Trends that shift — Less formal than ARIMA
4. ARIMA / SARIMA — Auto-Regressive Integrated Moving Average: formal time-series modeling of correlations, seasonality, trends — Formal statistical rigor — Needs care with stationarity and parameter selection
5. Prophet (Facebook) — Built for business forecasting at scale: holidays, multiple seasonalities, trend changes — "Better than linear without learning ARIMA"; the industry-standard middle ground — Tuning effort
6. ML (LSTM, Transformer, custom) — Complex patterns, many inputs, large datasets — Fraud detection, recommendation, retail demand forecasting — Heavyweight; almost always overkill for "how many servers next month"

The 80/20 of forecasting for SRE

Linear with seasonality — 80%
Prophet — 15%: irregular events, complex seasonality
ML — 5%: unusual patterns, business-critical accuracy

Accuracy metrics

MAPE (Mean Absolute Percentage Error): average of |actual - forecast| / actual. The standard metric.
MAE (Mean Absolute Error): average of |actual - forecast|, in absolute units.
Forecast bias: mean(forecast - actual). Persistently positive means consistent over-forecasting.

> [!TIP] For capacity planning, aim for MAPE [!NOTE] Driver-based forecasts beat naive trends when leading indicators are good, and lose when the business indicators are themselves uncertain.

Capacity-specific gotchas

Forecasting averages — Misses peaks: forecast peak QPS, not average
Compositional changes — A new feature changes the mix of work; old QPS forecasts undercount its cost
Failure-mode shifts — Onboarding an APAC customer moves peak from 8pm to 6am; region-level forecasts matter
Non-stationarity — Past patterns may not hold; retrain after a major product change

Key concepts

Linear with seasonality: 80% of SRE capacity needs. Daily/weekly/yearly cycles + linear growth trend.
Prophet (Facebook): Open-source forecasting for business time series. Handles holidays, multiple seasonalities. Industry standard middle-ground.
MAPE: Mean Absolute Percentage Error. Standard accuracy metric for forecasts. < 20% is good for quarterly capacity.
Driver-based forecast: Predict from leading indicators (MAU, marketing spend, business projections), not from past traffic alone.
Forecast bias: Persistent over- or under-forecasting. mean(forecast - actual) over a window. Should trend to 0.

When to use it

Quarterly capacity planning — model the next 90 days
Pre-event sizing — Black Friday, marketing campaign
Post-launch capacity tracking — actual vs forecast for new features
Cost forecasting — predict cloud spend

Interview questions

When should I use Prophet vs simple linear forecasting?
How do you incorporate marketing campaigns or product launches into a forecast?
What's a good MAPE for capacity-planning forecasts?