A Cheat-Sheet for Prompt Engineering for Time Series

This is an explainer on how to talk to an AI so it can make better predictions about numbers that change over time, like daily sales or website visits.

Big idea

Treat the AI like a smart but inexperienced intern: it knows a lot of facts, but nothing about your specific business or calendar unless you explain it clearly. If you give it the right story, structure, and tasks, it can become a helpful junior forecaster instead of just guessing from raw data.

Step 1: Tell the story first

Before asking for predictions, describe how your numbers usually behave: for example, “Saturdays are busier, December is strong, end-of-month paydays cause spikes.” These simple sentences give the AI a “feel” for your business rhythm so its predictions are grounded in reality.

Step 2: Ask it to summarize the data

Instead of jumping straight to “What will happen next?”, first ask the AI to summarize the past: average levels, how much they vary, and any unusual days that stand out. This forces the AI to notice patterns and odd points, which it can then use to justify and improve its forecast.

Step 3: Combine AI with classic models

Do not expect the AI to do all the heavy maths on its own. Let traditional forecasting tools (like standard statistical models) handle the number crunching, and use the AI to add context, scenarios, and explanations around those forecasts.

Step 4: Use clean, structured data

Feed the AI data in a tidy, consistent format, such as a simple list of records with “date, value, and notes/metadata” instead of messy spreadsheets. Clear structure makes it much easier for the AI to understand and refer back to different parts of the time series.

Step 5: Be specific in what you ask

Break your request into clear pieces: for example, “Give me a 7day forecast, a 30day forecast, and a version assuming a marketing promotion, each with uncertainty ranges.” When you spell out the different scenarios you care about, the AI can focus on choosing and explaining numbers instead of guessing the task.​

Step 6: Separate detection and explanation

For spotting anomalies, first use simple rules or tools to flag suspicious days (for example, days far above or below normal). Then ask the AI to explain what might have happened on those days and whether they deserve attention, turning raw spikes into understandable stories.

Step 7: Add your business knowledge

Finally, always mix in your domain knowledge: pay cycles, holidays that move, stockouts, promotions, or supply issues that affect your numbers. When you include these details in the prompt, the AI’s analysis starts to sound less generic and more like a knowledgeable manager who understands your business.