Every model extends a pattern
A forecasting model looks at past values, finds a pattern it can describe, and extends that pattern into the future with a range around it. What changes between models is the kind of pattern they look for and how many series they learn from.
| Family | How it extends the past | Example | What the evidence shows |
|---|---|---|---|
| Simple benchmarks | Repeat the last value (naive), the value from the same season last year, the historical average, or the average change (drift) | Naive, seasonal naive, mean, drift | Described as "extremely simple and surprisingly effective" and used as benchmarks throughout Forecasting: Principles and Practice, 5.2 |
| Exponential smoothing | Weighted averages of past values, with weights that shrink as observations get older (FPP3, chapter 8); variants add trend and season | Simple exponential smoothing, Holt, damped trend | The M4 Comb benchmark, an average of these three, ranked 19th with OWA 0.898 (official M4 scores) |
| ARIMA | Describes the autocorrelations in the data: how each value relates to earlier ones | ARIMA, seasonal ARIMA | With exponential smoothing, one of "the two most widely used approaches to time series forecasting" (FPP3, chapter 9) |
| Combinations | Average the forecasts of several methods | Montero-Manso, Talagala, Hyndman and Athanasopoulos; Pawlikowski, Chorowska and Yanchuk | Second and third in M4, OWA 0.838 and 0.841 |
| Machine learning | Learn the pattern from many series at once | M4's MLP and RNN benchmarks; Trotta's submission | Best pure machine learning submission ranked 23rd (OWA 0.915); the MLP and RNN benchmarks scored 1.642 and 1.482, worse than Naive2 |
| Hybrids | Combine a statistical model with a neural network | Smyl (Uber Technologies) | Won M4 with OWA 0.821, 8.6% better than Comb |
| Pretrained foundation models | Train one model on a large collection of public series, then forecast new series without retraining (zero-shot) | Chronos (Ansari et al., 2024), TimesFM (Das et al., 2024) | Chronos reports "comparable and occasionally superior zero-shot performance" on new datasets across a 42-dataset benchmark; TimesFM reports zero-shot accuracy that "comes close to" supervised models trained on each dataset |
Sources checked 13 November 2026. M4 figures come from the organisers' official scores file; the paper is Makridakis, Spiliotis and Assimakopoulos, "The M4 Competition: 100,000 time series and 61 forecasting methods", International Journal of Forecasting 36(1), 2020.
What the M4 Competition found
The M4 Competition, run by the University of Nicosia from 1 January to 31 May 2018, asked teams to forecast 100,000 real series, from yearly to hourly (organiser's page). Each method was scored with sMAPE and MASE, and ranked by OWA, the average of the two after dividing each by the Naive2 benchmark's score.
| Rank | Method | Type | sMAPE | MASE | OWA |
|---|---|---|---|---|---|
| 1 | Smyl | Hybrid | 11.374 | 1.536 | 0.821 |
| 2 | Montero-Manso, Talagala, Hyndman and Athanasopoulos | Combination (statistical and ML) | 11.720 | 1.551 | 0.838 |
| 3 | Pawlikowski, Chorowska and Yanchuk | Combination (statistical) | 11.845 | 1.547 | 0.841 |
| 8 | Legaki and Koutsouri | Statistical | 11.986 | 1.601 | 0.861 |
| 18 | Theta (benchmark) | Statistical | 12.309 | 1.696 | 0.897 |
| 19 | Comb (benchmark: SES, Holt, damped) | Combination (statistical) | 12.555 | 1.663 | 0.898 |
| 23 | Trotta | Machine learning | 12.894 | 1.682 | 0.915 |
| 35 | Naive2 (benchmark) | Statistical | 13.564 | 1.912 | 1.000 |
| 39 | Naive (benchmark) | Statistical | 14.208 | 2.044 | 1.058 |
| 55 | RNN (benchmark) | Machine learning | 21.152 | 2.685 | 1.482 |
| 57 | MLP (benchmark) | Machine learning | 21.653 | 3.225 | 1.642 |
Source: "Evaluation and Ranks.xlsx" in the organisers' M4-methods repository, checked 13 November 2026. Lower is better in every column. Naive2 is the naive forecast applied to seasonally adjusted data, as defined in the organisers' benchmark code.
Three things stand out. The top three entries all mixed methods. Seventeen entries beat the Comb benchmark, a plain average of three textbook methods, and everything below rank 19 did not. And the two machine learning benchmarks finished 55th and 57th, behind both naive benchmarks.
What none of these models do
A model only sees the data it is given. Hyndman and Athanasopoulos note that prediction intervals from ARIMA models assume "the historical patterns that have been modelled will continue into the forecast period" (FPP3, 9.8), and the same holds for every family above. A pretrained model has seen more series, not the future: when a pattern breaks, every method in the table is extending the old one.
A chat assistant that answers questions about future events, such as elections or prices, is doing something different again: judgemental forecasting from text. How those assistants score against human forecasters is on where forecasts are reliable.
Sources
Checked 13 November 2026:
- Hyndman, R. J. and Athanasopoulos, G., Forecasting: Principles and Practice, 3rd ed., OTexts: 5.2 Some simple forecasting methods, 8 Exponential smoothing, 9 ARIMA models and 9.8 Forecasting
- M4 Competition organisers, M4-methods repository (official scores and benchmark code) and competition page, University of Nicosia
- Makridakis, S., Spiliotis, E. and Assimakopoulos, V. (2020), The M4 Competition: 100,000 time series and 61 forecasting methods, International Journal of Forecasting 36(1), 54-74
- Ansari, A. F. et al., Chronos: Learning the Language of Time Series, arXiv, revised November 2024
- Das, A., Kong, W., Sen, R. and Zhou, Y., A decoder-only foundation model for time-series forecasting, arXiv, revised April 2024
Next step
Before trusting any model, compare it with the naive forecast on your own data using the forecast score calculator.
Frequently asked questions
- What is the simplest forecasting method?
- The naive method, which sets every forecast to the last observed value. Hyndman and Athanasopoulos note it works remarkably well for many economic and financial time series, and it is the standard benchmark a forecasting model has to beat.
- Did machine learning win the M4 Competition?
- Not on its own. The winner was a hybrid of statistical and machine learning methods (OWA 0.821). The best pure machine learning submission scored 0.915, behind the Comb benchmark, a plain average of three exponential smoothing methods (0.898).
- What does OWA mean in the M4 results?
- Overall weighted average. It averages a method's sMAPE and MASE, each divided by the score of the Naive2 benchmark, so Naive2 scores 1.000 and lower is better.