Specialist AI: forecasting, graphs and active learning · Module 1: Specialist AI: forecasting, graphs and active learning in practice
Time-series forecasting and temporal leakage
Learning outcome: Evaluate forecasts using only information available at the forecast date.
Forecasting predicts future values from information available at a defined time. Randomly mixing past and future observations into training and test sets can create an unrealistic evaluation. Seasonality, trend and changing operating conditions also affect what a forecast can learn.
Begin with a simple baseline such as repeating the last observed value or the comparable previous seasonal value. Evaluate on later periods and use multiple forecast origins where appropriate. Keep the horizon explicit: next day and next quarter are different tasks. Do not treat an explanatory narrative from a language model as a validated forecasting method.
Worked example
A fictional café predicts tomorrow's demand. Its dataset includes tomorrow's actual staffing adjustment, decided after the morning rush begins. That field would not be available when the forecast is needed. A valid test removes it and compares the model with a simple same-weekday baseline using earlier weeks only.
Try it yourself
Design a four-week evaluation timeline for a next-day forecast. Mark the information cut-off, forecast horizon and a baseline. Include a day with a special event.
You can do this exercise on paper. If using an assistant, keep to the fictional material and an approved tool.
Compare your approach
The special event can be an input only if its information was known before the cut-off. Report how the model handles unusual days separately from ordinary ones. A future observation must not leak into model preparation.
Check your understanding
Which information can a next-day forecast legitimately use?
- Information available at the stated forecast cut-off
- Tomorrow’s realised demand
- A field populated after tomorrow’s service ends
Show the answer and explanation
Answer: Information available at the stated forecast cut-off. The forecast must use information that exists at its cut-off. Realised demand and fields populated after service begins leak future information into the evaluation.
Workbook
Download the course workbook to record your answer and compare it after the knowledge check.