Time Series
Advanced 15 min read Time Series

Modern Forecasting: Prophet to Gradient Boosting

When 'forecast by curve + features beats fit a lot to history'

Beyond classical models, modern forecasters fit simple-but-flexible curves (Prophet-style), rich-calendar GBDT feature models, or deep sequence models — each with the same discipline of clean windows and honest backtests.

Prophetgradient boostingforecastcalendar featuresdeep forecastinghierarchical
Prerequisites: Lag Features and Windowing
Before we start

What you will learn

  • Position Prophet, GBDT feature models and DeepAR/Transformer styles
  • Choose by horizon, series count and data volume
  • Keep classical rigor (windows, baselines) for every modern model
Introduction

What is this?

Modern forecasting leans on the same decomposition as classical stats but fits it with flexible machinery: Prophet fits an interpretable curve (trend + seasonality + holiday regressors) with robust smoothing; gradient boosting consumes lag/rolling/calendar features for tabular tabular-native learners; deep models (DeepAR, N-BEATS, temporal transformers) learn across many series at once. They trade statistical parsimony for flexibility — and inherit all the windowing discipline required to be honest.

Motivation

Why do you need it?

  • Holiday-aware, multi-series, or high-frequency calendars defeat SARIMA's rigid assumptions.
  • GBDTs turn your available features (price, promo, weather) into forecast gains classical P(D,Q) can't express.
  • Cross-series deep models exploit shared structure when hundreds of related series exist.
  • An interpretable Prophet can be the right 'good enough' for capacity planning.
Think about it

Mental model

Imagine

A mechanic with a full toolbox vs a specialist lathe

Think of

Classical vs modern forecasters

The specialist (ARIMA) is precise on its material (stationary-ish single series). The mechanic (feature ML) adapts to whatever the job needs — price moves, holidays, external shocks — as long as the measurements (features) are sound. You bring the specialist when the material is known, the mechanic when the job is complicated.

Mechanics

How it works

  • Prophet: fit a Bayesian-ish trend (piecewise linear/logistic growth), Fourier seasonal terms, holiday/regressor tables; forecast with posterior intervals.
  • GBDT forecast: build lag/rolling/calendar matrix, train LightGBM/XGBoost per horizon (or recursive), walk features forward.
  • Deep: embed time + series id, train on many series with rolling windows, sample forecast paths for intervals.
  • Common to all: fixed as-of features, future-only validation, naive baselines, per-horizon scoring.
  • Hierarchical reconciliation sums coherently across levels (SKU → category → region).

Choosing among forecast families.

One series, smooth
Holidays + trends
External features
Many series
Validate
Real world

In production

  • A supply-chain team Prophet-forecasts capacity needs monthly — holidays modeled explicitly, no maintenance.
  • A pricing team feeds price changes into LightGBM with 90-day lags, winning 8% MASE over ARIMA at h=14.
  • A platform with 5,000 store demand series trains a single deep forecast model learning shared weekly patterns.
Guidelines

Best practices

  • Stay honest: every modern model must clear the same backtest + baseline gates as ARIMA.
  • Use holiday calendars (national + regional) as first-class features in Prophet/GBDT.
  • Watch for multi-collinearity between lags and calendar in GBDT — mitigation rarely needed, results are the verdict.
  • For deep/GBDT, deduplicate horizons and walk-forward refits; interval coverage needs residues, not guesses.
Watch out

Common mistakes

  • Assuming 'modern' means 'no leakage rules' — rolling windows computed full-history still cheat.
  • Fitting one GBDT for all horizons and reporting the h=1 number.
  • Prophet on series with double seasonality (hourly + weekly) without the proper Fourier terms.
  • Comparing each family with a DIFFERENT backtest window — the leaderboard is only fair on identical runs.
Root cause analysis

Common issues

Error
GBDT forecast 'perfect' then collapses after a promo window
Root cause

Features included promo without a modeling regressor for its effect

Fix

Add promo/price as explicit features and validate on a promo-excluded backtest month.

Error
Deep forecast's intervals are far too narrow
Root cause

Sampling doesn't include model/parameter uncertainty

Fix

Widen via residual-based ensemble or quantile training (pinball loss) rather than single-sample paths.

When things break

Debugging tips

  • Plot one full backtest series next to the truth; the aggregate metric can't show a failing month.
  • If holiday flags lift performance but the calendar table is wrong, results are untrustworthy — fix the calendar first.
  • For GBDT, importance lists that top with lags means calendar may be under-used; holiday-only models underperform.
  • Prophet flatlines trend? Model under-smooths season; increase changepoint regularization.
Interviews

Interview checkpoint

Test yourself

Quick quiz

0/2
1

Deep cross-series forecasters earn their place when:

2

Modern forecast models are exempt from as-of feature discipline.

Go deeper

Official documentation

Prophet documentation
docs
GluonTS deep forecasting
docs