...forecast time series
Jump to navigation
Jump to search
| AI Solver map |
|---|
|
Your current page is in bold
0 · Start
1 · Predict & analyze
tables, numbers, events
Predict values
Predict categories Forecast time series Cluster / segment Find outliers Make recommendations Find key drivers 2 · Understand unstructured data
text, documents, images, audio
3 · Generate & converse
4 · Decide & act
5 · Build it right
applies to every path
|
YouTube ... Quora ...Google search ...Google News ...Bing News
- AI Solver ... Forecasting ... ...predict values ... ...find outliers ... ...optimize decisions ... Excel - Forecasting
Use this page when you predict future values of something measured over time: sales per store, energy load, web traffic, river levels. Forecasting is regression with a time order, so the rules for testing are different (never shuffle time).
Start here
First, always: build a seasonal-naive baseline (next Monday = last Monday). Every model must beat it.
- How many series, and how much history?
- One or a few series with clear trend/seasonality → statistical models: exponential smoothing (ETS) and ARIMA (e.g., the statsforecast library); in a spreadsheet, Excel - Forecasting (FORECAST.ETS). Business data with holidays → Prophet
- Many related series (thousands of products or stores) → one global model trained across all series: LightGBM with lag and calendar features, or deep models (N-BEATS/N-HiTS, Temporal Fusion Transformer, PatchTST)
- Short history, new products, or you need a strong baseline today → time-series foundation models, zero-shot: Chronos-2 (Amazon), TimesFM (Google), Moirai (Salesforce), TimeGPT (Nixtla, as an API), Toto (Datadog, for observability metrics). Fine-tune if you have data.
- Do outside drivers matter (price, promotions, weather, holidays)?
- Yes → models that take covariates: gradient boosting with features, Temporal Fusion Transformer, Chronos-2
- Do decisions need a range (stock levels, staffing, risk)?
- Yes → probabilistic forecasts: quantiles from Chronos-style models, DeepAR, or quantile gradient boosting; or conformal intervals
- Is it a physical system (weather, rivers, power grids)?
- Yes → specialized physical or hybrid AI models (e.g., GraphCast for weather, flood models), not generic sales forecasters
- Want to flag unusual values rather than predict them? → ...find outliers
- Will the forecast drive an order, schedule or price? → forecast, then optimize: ...optimize decisions
Quick picks
| Situation | Try first | Then compare |
|---|---|---|
| One series, clear seasonality | ETS / ARIMA | Prophet, a foundation model zero-shot |
| Thousands of related series | Global LightGBM with lags and calendar features | N-HiTS, TFT, Chronos-2 / TimesFM |
| Little history or cold start | Chronos-2 or TimesFM zero-shot | Hierarchical / grouped statistical models |
| Strong external drivers | Gradient boosting with covariates | TFT, Chronos-2 with covariates |
| Need intervals for decisions | Quantile forecasts | Conformal prediction around your best model |
| Spreadsheet users | Excel FORECAST.ETS | Prophet |
Real-world examples
- Retail demand (M5 competition) — forecasting Walmart unit sales across 42,840 hierarchical series; most top entries used LightGBM global models, and simple statistical methods remained hard-to-beat benchmarks. M5 Forecasting - Accuracy | Kaggle
- Flood warnings — Google's AI river-flood forecasts give up to 7 days of warning in more than 100 countries, covering about 700 million people. Flood Forecasting | Google Research
- Weather — DeepMind's GraphCast (a graph neural network) produces 10-day global forecasts in under a minute and beat the leading operational system (ECMWF HRES) on 90% of 1,380 test targets. Learning skillful medium-range global weather forecasting | Science
- Cautionary tale — Zillow closed its home-buying business in 2021 after its price forecasts failed in a fast-moving market. See ...predict values.
What's new
- Foundation models turn forecasting into model selection instead of model training. On the public GIFT-Eval benchmark, Chronos-2 reported better accuracy than earlier leaders such as TimesFM-2.5 and TiRex, and it supports covariates and multiple related series. Chronos-2: From Univariate to Universal Forecasting | arXiv
- Read leaderboards carefully: some models were pretrained on data overlapping benchmark training sets, so test on your own history.
Watch out
- Leakage: features that use future information (e.g., the month's final total) inflate test scores.
- Wrong validation: use rolling-origin backtests (train on the past, test on the next window, repeat), not random splits.
- Regime changes (pandemics, price shocks) break history; add judgment and scenario planning.
Measure success
MASE (vs. naive), WAPE or sMAPE (relative error), quantile loss or CRPS (for ranges), bias (systematic over/under-forecasting), all averaged over several backtest windows. More: ...evaluate and trust AI
Learn more
- Forecasting: Principles and Practice (free book) | R. J. Hyndman & G. Athanasopoulos
- statsforecast: fast statistical forecasting | Nixtla
- Chronos forecasting models | Amazon Science
- TimesFM | Google Research
- Prophet | Meta
- GIFT-Eval forecasting leaderboard | Salesforce
- Forecasting ... Excel - Forecasting
Where next? ◀ Predict & analyze · ...predict values · ...find outliers · ...optimize decisions · ▶ ...evaluate and trust AI