01

Pick a scenario

Each scenario has different trend, seasonality, and noise characteristics. Switching between them changes the whole interactive section below — same code, different data shape.

02

Anatomy of a time series

Almost any operational series decomposes into three components: trend (long-term direction), seasonality (repeating patterns at fixed intervals) and residuals / noise (what’s left after removing trend and seasonality). Most forecast methods differ in how explicitly they model each piece.

Key concept

Decomposition is not just diagnostic — it determines which forecasting family to reach for. Strong seasonal patterns? Use Holt-Winters or SARIMA. Pure trend with little seasonality? A simpler model often wins on short horizons.
03

Forecast playground

Forecasting grant disbursements (M€) with Holt-Winters over a 6-month horizon. Switch methods, drag the horizon, or change scenarios above to see how the forecast adapts.

Method

Holt-Winters

Triple exponential smoothing, explicitly models level, trend and seasonality.

Tap for full details

Grant disbursements — 24 months history + 6-month forecast (Holt-Winters)

Reading the grant disbursements chart

Look at the Oct–Dec values, disbursements double the rest-of-year average as countries close out their grant cycles. The baseline also drifts upward as the portfolio grows.
04

Which method actually wins?

One forecast on one slice of data isn’t enough — you need a backtest. Each method runs on the first 24 months and predicts the last 6; we measure the error with MAPE (mean absolute percentage error). Lower is better.

Head-to-head MAPE across methods — 6-month holdout backtest

On grant disbursements: winner

Prophet-style wins with 3.8% MAPE. Naïve and moving average miss the Q4 spike entirely. Holt-Winters and Prophet recover both the trend and the seasonal pattern.
05

Connecting back to EVM

This stack plugs straight into Earned Value Management: the Estimate at Completion (EAC) is a forecast of total project cost, and its quality depends on how well we predict the future actual-cost (AC) curve.

The naïve EAC formula (AC + (BAC − EV) / CPI) treats future cost performance as a constant. Holt-Winters or Prophet can replace that constant with a properly modeled trajectory, and the backtest above tells you exactly how much the upgrade is worth on your specific data.

Takeaway

Don’t treat EAC as a static ratio — model the AC curve forward with the best method from the backtest and you’ll catch schedule pressure weeks earlier.
06

To go further

Worth exploring next: ARIMA / SARIMA (via pmdarima.auto_arima) for series where Holt-Winters plateaus; rolling-origin cross-validation instead of a single holdout; MAPE vs. RMSE depending on audience; hierarchical forecasting for country-level and portfolio-level reconciliation; and modern neural approaches like NeuralProphet or N-BEATS when the structure is genuinely non-linear.

Rule of thumb

Always start with the simplest method that fits the data shape. Complexity is a cost — pay it only when the backtest shows it buys you meaningful accuracy.