Back to library

📈Know When Not to Use ML for Time-Series Forecasting

Stop reaching for LSTMs on tiny series — enforce the baseline ladder (naive, seasonal-naive, ARIMA, then ML), backtest each one properly on real data, and write the decision criteria your team will use to escalate to ML only when the simple model is genuinely beaten.

Applied14 drops~2-week path · 5–8 min/daytechnology

Phase 1Why the Simple Forecast Usually Wins

Why naive forecasts beat fancy models so often

4 drops
  1. Yesterday's value is the model to beat

    6 min

    Yesterday's value is the model to beat

  2. There is no universally best forecasting model

    6 min

    There is no universally best forecasting model

  3. 200 points is not enough for a transformer

    6 min

    200 points is not enough for a transformer

  4. Pure trend is not what ML is for

    5 min

    Pure trend is not what ML is for

Phase 2Run the Baseline Ladder on One Real Series

Backtest naive, seasonal-naive, ARIMA, and an ML model

5 drops
  1. Choose the series you'll benchmark for the rest of this path

    5 min

    Choose the series you'll benchmark for the rest of this path

  2. Score the two zero-parameter models first

    8 min

    Score the two zero-parameter models first

  3. Add ARIMA — the strongest classical option

    10 min

    Add ARIMA — the strongest classical option

  4. Now train the ML model — and don't tilt the playing field

    12 min

    Now train the ML model — and don't tilt the playing field

  5. Read your scoreboard like an engineer, not a fan

    8 min

    Read your scoreboard like an engineer, not a fan

Phase 3When Classical Methods Beat ML by Design

When short series and pure trend favor classical methods

4 drops
  1. Why ML can't escape overfit on short series

    8 min

    Why ML can't escape overfit on short series

  2. When the signal-to-noise ratio is too low for any model

    7 min

    When the signal-to-noise ratio is too low for any model

  3. ML's win must be big enough to pay for itself

    8 min

    ML's win must be big enough to pay for itself

  4. The four situations where ML really is the right call

    8 min

    The four situations where ML really is the right call

Phase 4Write Your Team's ML-Escalation Criteria

Write your team's ML-escalation decision criteria

1 drop
  1. Write your team's ML-escalation criteria

    20 min

    Write your team's ML-escalation criteria

Frequently asked questions

When should I use ML instead of ARIMA for time-series forecasting?
This is covered in the “Know When Not to Use ML for Time-Series Forecasting” learning path. Start with daily 5-minute micro-lessons that build from fundamentals to hands-on application.
Why does a naive forecast often beat a neural network on small series?
This is covered in the “Know When Not to Use ML for Time-Series Forecasting” learning path. Start with daily 5-minute micro-lessons that build from fundamentals to hands-on application.
What is the no-free-lunch theorem and how does it apply to forecasting?
This is covered in the “Know When Not to Use ML for Time-Series Forecasting” learning path. Start with daily 5-minute micro-lessons that build from fundamentals to hands-on application.
How do I properly backtest a forecasting model?
This is covered in the “Know When Not to Use ML for Time-Series Forecasting” learning path. Start with daily 5-minute micro-lessons that build from fundamentals to hands-on application.
What's the minimum series length before ML makes sense for forecasting?
This is covered in the “Know When Not to Use ML for Time-Series Forecasting” learning path. Start with daily 5-minute micro-lessons that build from fundamentals to hands-on application.