Learn · Beginner

Time-series forecasting: predicting the future without letting it leak into the past

Time-series forecasting predicts future observations using information that was available before a specified cutoff. It matters because the order of observations changes what a model is allowed to know: a method that appears accurate after seeing the future can be useless when a real decision must be made today.

A time series is a sequence of measurements indexed by time: hourly electricity use, daily sales, weekly hospital admissions or minute-by-minute traffic. Forecasting asks for one or several future measurements. The distance from the cutoff to the prediction is the horizon. Predicting demand one hour ahead and predicting it a year ahead are different tasks even when they use the same historical chart.

Imagine running a bakery. You must choose tonight how much bread to prepare for tomorrow. Yesterday’s sales, the weekday and a weather forecast published this afternoon may be available. Tomorrow’s actual temperature and tomorrow’s final sales are not. A valid experiment reconstructs that information boundary instead of allowing the model to use everything that eventually ended up in a spreadsheet.

Three recurring patterns help organize the problem. A trend is a longer-term change, such as steadily rising demand. Seasonality is a repeating structure, such as a Saturday rush or an annual holiday cycle. Noise is variation the available information does not reliably predict. These categories are useful descriptions, not guarantees that every series cleanly separates into three independent parts.

The first competitor should often be embarrassingly simple. Predict that the next observation equals the latest one, or that next Monday looks like last Monday. These are persistence and seasonal-naive forecasts. A complicated model that cannot improve on the relevant baseline may add cost without useful information. Comparing against a weak baseline can make ordinary pattern matching look like a breakthrough.

Classical forecasting uses tools such as autoregression, which predicts values from earlier values, and exponential smoothing, which updates estimates of level and seasonality as observations arrive. Modern neural methods can learn across many related series, incorporate additional variables or produce several future steps together. The architectural choice is secondary to whether the model is evaluated on the decision it must actually support.

Recent foundation models extend that idea across datasets. In A decoder-only foundation model for time-series forecasting, Abhimanyu Das and colleagues describe pretraining for forecasting across time series. Abdul Fatir Ansari and colleagues’ Chronos maps numerical sequences into tokens and trains a language-model-style predictor. These approaches transfer statistical experience across series; they do not turn every changing system into a reliably predictable one.

The Timestral announcement in today’s dossier is a current product example of this broader direction. Mistral describes a family of forecasting models, including CPU execution. That announcement does not remove the need to compare a model on the user’s own horizon, available variables and data distribution. A public ranking measures a specified experiment, not universal suitability.

Evaluation must move forward through time. In a rolling-origin test, train or condition on an initial period, forecast the next interval, advance the cutoff and repeat. The model never receives observations that had not arrived at that cutoff. This is a forecasting-specific application of holdout evaluation, with chronology doing work that a random row split cannot.

Leakage can hide in preprocessing. Computing a normalization scale from the entire dataset exposes future information. Filling an old missing value with a later observation may do the same. A feature labeled as a forecast may actually be a revised historical value published after the decision date. The right question is not merely when an event occurred, but when the forecasting system could have known about it.

A point forecast also leaves out uncertainty. A bakery may want a central estimate for staffing and a high-demand bound for avoiding stockouts. A probabilistic forecast describes a range or distribution of possible outcomes. Its usefulness depends on calibration: events described as having a given probability should occur at roughly that frequency under comparable conditions.

Finally, prediction differs from intervention. Sales falling after a price change does not prove the price caused the decline; weather or competitors may have changed too. Distribution shift can also break patterns that held historically. A reliable forecasting workflow preserves information timing, beats a meaningful baseline, measures uncertainty and keeps checking whether the world still resembles the one the model learned.

Ground Truth. (2026, October 16). Time-series forecasting: predicting the future without letting it leak into the past. Ground Truth. https://groundtruth.day/learn/time-series-forecasting.html

@misc{groundtruth:time-series-forecasting,
  title  = {Time-series forecasting: predicting the future without letting it leak into the past},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {oct},
  url    = {https://groundtruth.day/learn/time-series-forecasting.html}
}

Topics: forecasting · time-series · fundamentals · evaluation