# AI forecasting vs. traditional forecasting: when does AI actually win?

> AI forecasting beats traditional methods in weather and large retail data, but not everywhere. What the M4 and M5 competitions show, and when to use which.

Oct 14, 2026 · AI forecasting · Sikt Intelligence · https://www.siktintelligence.com/blog/ai-forecasting-vs-traditional-forecasting

**Short answer: AI wins when there is a lot of related data, and simple methods still win surprisingly often when there is not.** In weather, AI models now beat the leading conventional forecasts on most targets. In large retail datasets, machine learning won the M5 competition outright. But on short, single time series, classic statistical methods remain hard to beat, and on real-world events the best AI is close to, not clearly ahead of, the best human forecasters. Here is what the evidence shows, area by area, and how to tell which approach fits your problem.

## What counts as "traditional" forecasting

Traditional forecasting covers two things:

- **Statistical methods** with a fixed form: moving averages, exponential smoothing, ARIMA and regression. They assume the future looks like a smoothed version of the past, plus trend and season. They are fast, cheap and easy to explain.
- **Human judgment:** analysts, expert panels and polls of experts. For one-off events, such as an election or a central bank decision, judgment was long the only option.

## What counts as AI forecasting

[AI forecasting](/blog/ai-forecasting-explained) learns the patterns from data instead of assuming them:

- **Machine learning** such as gradient-boosted trees (LightGBM, XGBoost), trained on many related series at once.
- **Neural networks** built for time series, and since 2024 **time-series foundation models** such as Google's TimesFM and Amazon's Chronos, which are pretrained on huge collections of series and can forecast a new one without being trained on it ([Das et al.](https://arxiv.org/abs/2310.10688), [Ansari et al.](https://arxiv.org/abs/2403.07815)).
- **Language-model forecasters** for events, which read the news, start from a [base rate](/blog/base-rates-forecasting) and give a probability. A well-built one is often called an [AI superforecaster](/blog/what-is-an-ai-superforecaster).

## The best evidence: forecasting competitions

Claims that "AI beats traditional forecasting" are easy to make and hard to check. The most trusted evidence comes from open competitions, where every method forecasts the same data and is scored on results nobody had seen. The best known are the M competitions, led by Spyros Makridakis since 1982.

| | M4 (2018) | M5 (2020) |
|---|---|---|
| Data | 100,000 time series from many fields | 42,840 series of Walmart unit sales |
| Winner | A hybrid of exponential smoothing and a neural network | Machine learning, built on LightGBM |
| Pure machine learning | Mostly worse than simple statistical benchmarks | Every top entry, ahead of all statistical benchmarks |

The two results look contradictory, but they fit together ([Makridakis et al., 2020](https://doi.org/10.1016/j.ijforecast.2019.04.014); [Makridakis et al., 2022](https://doi.org/10.1016/j.ijforecast.2021.11.013)):

- **M4's series were mostly separate and short**, from different fields. There was little shared structure for a learning method to exploit, and simple methods, and combinations of them, did very well.
- **M5's series were related**: the same products, in the same stores, with prices, promotions and holidays. Machine learning could learn from all of them at once, and that is where it pulled clearly ahead.

> The question is not "AI or statistics?" but "how much related data is there to learn from?"

## Weather: the clearest win for AI

Weather is where AI forecasting has won most convincingly. In 2023, Google DeepMind's GraphCast was more accurate than the leading conventional medium-range forecast on 90% of 1,380 test targets ([Lam et al.](https://arxiv.org/abs/2212.12794)). In 2024, its probabilistic successor GenCast beat the world's top operational ensemble on 97.2% of 1,320 targets ([Price et al.](https://www.nature.com/articles/s41586-024-08252-9)).

Weather suits AI for the same reason M5 did: decades of consistent, global data, and the same kind of question asked millions of times. The full story is in [can AI predict the weather?](/blog/can-ai-predict-the-weather)

## Real-world events: AI against human judgment

For one-off events there is no long history of the same question, so the comparison is with human judgment. Here AI has moved from near chance to close to the best humans in about two years:

- On ForecastBench, several AI systems were "statistically indistinguishable from superforecaster-level accuracy" in July 2026, with caveats about the age of the human data ([Forecasting Research Institute](https://forecastingresearch.substack.com/p/ai-models-have-likely-reached-parity)).
- In Metaculus's spring 2026 tournament, professional forecasters still beat the top bots, but by a margin too small to be statistically significant ([Metaculus](https://www.lesswrong.com/posts/wZBbDqzfBjYG58CxK/futureeval-spring-results-pros-beat-bots-but-the-gap-is)).

So on events, AI is roughly level with the best of traditional judgment, and far cheaper and faster. We weigh both sides in [AI vs. superforecasters](/blog/ai-vs-superforecasters).

## Stock prices: where neither wins easily

Prices are the exception to everything above. A pattern that predicts returns gets traded away once enough investors use it, whether it was found by a statistician or a neural network. Research finds language models can read news and predict a stock's first reaction, but most of that move is not tradable, and the edge shrinks as AI use spreads. See [can AI predict the stock market?](/blog/can-ai-predict-the-stock-market) and our own [one-week stock market test](/blog/ai-stock-market-test).

## When to use which

| Situation | Better bet | Why |
|---|---|---|
| Thousands of related series, with prices, promotions or other drivers | AI (machine learning) | It learns shared patterns across series |
| A few short series with little extra data | Simple statistical methods | Too little data for AI to learn from; easy to explain |
| A brand-new product with no history | Judgment, or analogies with similar products | No model can learn from data that does not exist |
| Weather | AI weather models | Decades of global data; AI now leads on most targets |
| One-off events: elections, policy, conflicts | AI forecaster, checked against [prediction markets](/blog/what-is-a-prediction-market) | Close to the best human forecasters, at a fraction of the cost |
| Short-term stock prices | Neither, with any confidence | Edges are traded away |

A practical rule: **always run the simple method too.** If an AI model cannot beat a seasonal naive forecast or exponential smoothing on your own data, it is not worth the extra cost and complexity.

## How to compare them fairly

1. **Test on data the model has never seen**, from after the end of its training data. A model that has seen the answers will look brilliant and fail in use. For language models this is called lookahead bias.
2. **Use a simple baseline.** For numbers, a seasonal naive forecast; for events, the [base rate](/blog/base-rates-forecasting) or the market's price.
3. **Use proper scores.** For numbers, scaled errors such as MASE, and CRPS for probabilistic forecasts. For events, the [Brier score](/blog/brier-score-explained) and a [calibration](/blog/forecast-calibration) check.
4. **Use enough cases.** One good quarter or one called election proves little.
5. **Count the cost.** A small accuracy gain can be outweighed by data, computing and maintenance costs.

## Where Sikt Intelligence fits

Sikt Intelligence works on the events side: rate decisions, elections, geopolitical flashpoints and technology milestones. It puts a probability on each one next to the market's price, and every forecast is graded when the answer arrives, misses included. You can follow it live on the [2026 midterms](/midterms). Sikt is in research; [join the waitlist](/#access) for early access.

## Key takeaways

- AI forecasting wins when there is a lot of related data: weather, and large retail datasets such as the M5 competition.
- On short, single series, simple statistical methods still match or beat AI, as the M4 competition showed.
- On real-world events, the best AI forecasters are close to the best humans in 2026, and much cheaper.
- Stock prices resist both approaches, because any edge gets traded away.
- Always compare against a simple baseline, on unseen data, with proper scores.

## FAQ

### Is AI forecasting better than traditional forecasting?

Sometimes. AI is clearly better in weather and on large sets of related series with rich data, as in the M5 competition. On short, single series, simple statistical methods are often just as accurate, and much cheaper and easier to explain.

### What is the difference between AI and statistical forecasting?

Statistical methods such as exponential smoothing and ARIMA assume a fixed form for the pattern. AI methods learn the pattern from data, often across many series at once, and can use many more inputs.

### What did the M5 competition show?

That on 42,840 related series of Walmart sales, machine-learning methods, mostly built on LightGBM, beat every statistical benchmark. In the earlier M4 competition, on 100,000 mostly unrelated series, pure machine learning mostly did worse than simple benchmarks.

### Can AI replace human forecasters?

For many routine number forecasts, it already has. For one-off events, the best AI forecasters are close to top human forecasters in 2026, and humans who use AI assistants forecast better than those who do not. The strongest setup today combines both.

## Sources

- Makridakis, Spiliotis and Assimakopoulos (2020): [The M4 Competition: 100,000 time series and 61 forecasting methods](https://doi.org/10.1016/j.ijforecast.2019.04.014), *International Journal of Forecasting*
- Makridakis, Spiliotis and Assimakopoulos (2022): [M5 accuracy competition: Results, findings, and conclusions](https://doi.org/10.1016/j.ijforecast.2021.11.013), *International Journal of Forecasting*
- Das et al. (2024): [A decoder-only foundation model for time-series forecasting](https://arxiv.org/abs/2310.10688) (TimesFM)
- Ansari et al. (2024): [Chronos: Learning the Language of Time Series](https://arxiv.org/abs/2403.07815)
- Lam et al. (2023): [GraphCast: Learning skillful medium-range global weather forecasting](https://arxiv.org/abs/2212.12794), *Science*
- Price et al. (2024): [Probabilistic weather forecasting with machine learning](https://www.nature.com/articles/s41586-024-08252-9), *Nature*
- Forecasting Research Institute: [AI models have likely reached parity with superforecasters](https://forecastingresearch.substack.com/p/ai-models-have-likely-reached-parity) (July 2026)
- Metaculus: [FutureEval spring 2026 tournament results](https://www.lesswrong.com/posts/wZBbDqzfBjYG58CxK/futureeval-spring-results-pros-beat-bots-but-the-gap-is) (September 2026)