How to think about "the methods"
There are dozens of forecasting methods, but they line up on a single ladder from simple to complex. Each higher rung can capture more structure - and asks for more data, more setup, and more skill in return. As a leader you do not need the math. You need to know what lives on each rung, when it is the right reach, and the one accountability question that keeps the whole thing honest: did it beat the baseline on a backtest?
The method ladder, rung by rung 10 min live
Read this from the bottom up. The lowest rung is the honest starting line every other method has to beat. Each rung above adds power - and a price in data and setup. The point of the picture is not to memorise the names; it is to feel that higher costs more and is not automatically better.
LiveThe bottom three rungs - baseline, smoothing, ARIMA4 min▶
- Baselines (naive, seasonal-naive). "Tomorrow equals today," or "this Tuesday equals last Tuesday." No training, no math. This is the honest starting line - on seasonal data it is shockingly hard to beat, which is exactly why every model is measured against it.
- Exponential smoothing / ETS. A weighted memory of the recent past, weighting fresh days more than old ones. Fast, robust, works on little data, and strong for short horizons. The reliable workhorse when you have one series and not much history.
- ARIMA / SARIMA. Models the correlations inside the series - how today relates to recent days and to the same day last season. More setup and expertise to fit well, and it shines on stable, structured series (the Box-Jenkins tradition).
LiveThe top three rungs - Prophet, ML, deep4 min▶
- Prophet. Breaks a series into trend + seasonality + holidays, with knobs an analyst can read and adjust. Designed at scale to be approachable for non-specialists (Taylor & Letham, 2018). Great when calendars and holidays drive the series.
- ML / gradient boosting. Reframes forecasting as regression: build features (last week's value, rolling averages, day-of-week, promotions) and let a model like XGBoost or LightGBM learn the mapping. Pays off when you have many related series and rich context.
- Deep models (DeepAR, N-BEATS, TFT). Neural networks that learn across many series at once, good for long horizons and large data. The heaviest rung - most data, most compute, most expertise - and cited here, not trained from scratch. They earn their place only when the data volume justifies them.
The rung is a business decision, not just a technical one. A deep model that needs a team to maintain and thousands of series to shine is the wrong reach for a single weekly demand line. Match the tool to the problem, not to the hype.
The central lesson: simple is hard to beat 5 min live
The M-competitions are the field's biggest open bake-offs, run over thousands of real series. Their repeated finding is humbling: more complex is not automatically more accurate. Simple methods, and combinations of a few methods, are stubbornly hard to beat. Complexity has to earn its place with evidence, not with a good story.
Why this matters to a leader. When a vendor or a team proposes a heavy model, the value question is not "is it sophisticated?" It is "does it beat the seasonal-naive baseline on a fair backtest, by enough to justify the cost to build and run it?" If nobody can show you that number, the sophistication is decoration.
Which rung fits your situation 6 min live
Three questions settle most choices: how much history do you have, how many related series, and how long is the horizon. Start at the lowest rung that meets the need and climb only when a backtest says the next rung earns its keep.
Self-studyWhy combinations often win2 min read▶
A recurring M-competition result is that averaging a few reasonable forecasts often beats any single one of them. The errors of different methods partly cancel out.
- You do not have to pick one winner. A blend of, say, ETS and a boosting model is a legitimate - often better - answer.
- It lowers risk. If one method degrades, the blend cushions the fall.
- Ask about it. "Did we try a simple combination?" is a fair challenge to a single-model pitch.
Read a model choice like a leader ★ 6 min · think it through
A vendor pitches a deep-learning forecaster for your weekly demand. Walk the ladder in your head before you nod.
Locate the pitch on the ladder. A deep net is the top rung - the heaviest reach in data, cost, and expertise.
Check the situation against the chooser. One weekly series with a couple of years of history is "little data, one series" - the bottom rungs, not the top.
Ask the accountability question. "Did it beat seasonal-naive and a simple ETS on a backtest, and by how much?"
Decide on evidence. If the heavy model does not clearly beat the cheap ones on a fair backtest, the right call is the lower rung - and the vendor just told you a lot about themselves.
Try it yourself - this week ◐ 20-30 min total
- Pick one real forecasting need in your org - demand, staffing, cash-flow, web traffic, tickets.
- Answer the three chooser questions for it: how much history, how many related series, how long a horizon.
- Name the rung you would start on, and write one sentence justifying it against the trade-off.
- Write the accountability question you would ask the team building it, in your own words.
- Optional: skim FPP3 chapter 1 (Hyndman & Athanasopoulos) - the open textbook this course follows - for the "what can be forecast" framing.
Sources covered (cited)
Taught from the forecasting canon and the published empirical evidence. This session covers:
Three questions before you go 🎯 ◐ 90 seconds
1 · What sits on the lowest rung of the method ladder?
The baseline is the honest starting line. On seasonal data "same weekday last week" is hard to beat, which is why every method is measured against it.
2 · What is the central lesson of the M-competitions?
Across thousands of series, simple methods and blends repeatedly hold their own. Complexity must earn its place with backtested accuracy, not with a good story.
3 · A vendor pitches a heavy model. What is the leader's one question?
No baseline comparison means no evidence. The backtest replays the model on past data as if it were live; the margin over the baseline is what justifies the cost.