Hello Model
← Model library

Time-Series Forecasting

Predict future values from history: demand, traffic, sales, load.

Build a plan for this →

Typical projects

Forecast weekly sales per storePredict website traffic next monthForecast energy demand per hour

Three ways to build it

Starter

Pretrained forecasting model (Chronos / TimesFM) or Prophet

Zero-shot foundation models forecast without training; Prophet is simple and explainable.

Best for: No or little data, or new to ML

chronos-forecasting · Prophet · pandas

Standard

StatsForecast (AutoETS/AutoARIMA) + LightGBM with lag features

Fast, reliable statistical models plus a tree model for extra signals.

Best for: Some labeled data and Python experience

StatsForecast · MLForecast · LightGBM

Advanced

Global deep models (N-HiTS / TFT) across many series

When you have thousands of related series, one global model learns shared patterns.

Best for: Lots of data and an experienced team

NeuralForecast · PyTorch · Ray

How success is measured

MAPE or MAE measured with backtesting over several past periods

The data you'll need

  • You need a timestamp column and the value, ideally 2+ full seasonal cycles (e.g. 2 years for yearly patterns).
  • Include known future events: holidays, promotions, price changes.
  • Keep the timestamps regular (fill gaps explicitly).

Labeling

No manual labeling needed — the future values in your history are the labels.

Preparing the data

  • Resample to a regular frequency
  • Fill or flag missing periods
  • Add calendar features (day of week, holiday)
  • Never shuffle — always split by time

Start with a baseline

A seasonal naive forecast: 'same as the same day last week'. Surprisingly hard to beat.

Evaluating the model

  • Rolling-origin backtesting
  • Compare against the seasonal naive baseline
  • Check prediction intervals actually contain ~80–90% of outcomes

Monitoring in production

  • Forecast error each period
  • Structural breaks (new store, pandemic-style shocks)
  • Retraining on a schedule (weekly/monthly)

Common pitfalls

  • Random train/test split leaks the future
  • Ignoring holidays and promotions

Example code

CodeQuick start
python
# pip install chronos-forecasting pandas torch
import pandas as pd, torch
from chronos import BaseChronosPipeline

df = pd.read_csv("sales.csv", parse_dates=["date"])
pipe = BaseChronosPipeline.from_pretrained("amazon/chronos-bolt-small")
quantiles, mean = pipe.predict_quantiles(
    context=torch.tensor(df["sales"].values), prediction_length=28,
    quantile_levels=[0.1, 0.5, 0.9])
print(mean)
CodeTrain your own model
python
# pip install statsforecast pandas
import pandas as pd
from statsforecast import StatsForecast
from statsforecast.models import AutoETS, SeasonalNaive

df = pd.read_csv("sales.csv", parse_dates=["date"])
df = df.rename(columns={"store_id": "unique_id", "date": "ds", "sales": "y"})

sf = StatsForecast(models=[AutoETS(season_length=7), SeasonalNaive(season_length=7)], freq="D")
cv = sf.cross_validation(df=df, h=28, n_windows=4)   # backtesting
print(cv.head())
forecast = sf.forecast(df=df, h=28)
forecast.to_csv("forecast.csv", index=False)

Ready to build one? Get a personalised plan →

Or read about Text Classification next.