ChatGPT for Forecasting: Assumptions, Scenarios, and Checks

ChatGPT for forecasting works best when you use it to clarify inputs, compare explicit scenarios, and explain a model whose calculations you can check. Give it historical data and a defined question rather than asking it to invent a prediction. A forecast should show what would have to be true for the result to happen.

Tools like fenbs keep the approved assumptions and later review findings outside the conversation, so your next forecast starts from the same decisions. That record matters when a plausible explanation survives longer than the evidence behind it.

How do you start a forecast with ChatGPT?

1. Define one outcome, its units, the time horizon, and the business decision it will inform. Weekly support requests and annual sales need different inputs.
2. Prepare authorized historical data with dates, missing periods, and known changes clearly labeled. Keep a dated source copy.
3. Ask ChatGPT to list assumptions and missing information before calculating. Separate observed facts from proposed explanations.
4. Choose a simple baseline and explicit alternative scenarios. Run or verify the calculations in your spreadsheet or statistical software.
5. Review errors against historical periods, have a person approve the planning choice, and record when the forecast should be revisited.

Start small enough that you can inspect the whole result. For example, a hypothetical repair shop might estimate next month’s bookings from its own weekly history. The useful question is how a staffing choice changes under different booking levels, not whether the assistant can sound certain about demand.

What inputs make the forecast worth discussing?

Write a short data dictionary: what each column means, its unit, its source, and whether a blank means zero or unknown. Identify cancellations, partial weeks, reporting changes, and events that make one period unlike another. Ask the assistant to repeat these definitions before it uses the numbers.

Time order matters. NIST’s introduction to time series analysis↗ (opens in a new tab) explains why patterns such as trends and seasonal variation need attention. Do not shuffle dated observations into an ordinary random sample and assume you have tested future performance.

Avoid giving the assistant an unexplained total and asking for the next total. Provide the history available at the forecast date, the definition of the outcome, and any genuinely known future inputs. Keep information learned afterward out of the test; otherwise you are giving the forecast an unfair advantage.

This is one practical workflow in our AI at work library; the broader ChatGPT for business guide covers where the work fits in an organization.

How should you ask for assumptions and scenarios?

Ask for an assumption table before a forecast narrative. Each assumption needs a proposed value or range, an evidence source, and an owner who can confirm it. A generated explanation is a hypothesis until someone checks it against actual records.

Goal: estimate weekly bookings for the next four weeks to discuss staffing.
Use only the attached approved history. First list gaps, units, and assumptions.
Propose a simple baseline and two clearly labeled scenarios.
Do not invent external demand data or probabilities.
Show formulas, calculations, and a test using earlier periods.
Separate calculated results, explanations to investigate, and decisions for a person.

Change a meaningful driver in each scenario. For the repair shop, compare ordinary capacity with a temporary reduction in available appointments. Avoid labeling arbitrary low and high numbers as statistically likely outcomes. If you assign probabilities, document where those probabilities came from and how they were evaluated.

Record the planning scenario a person chose and why in fenbs Decisions and rules, so another assistant does not quietly replace the approved direction.

How do you check the calculations and uncertainty?

OpenAI’s data analysis documentation↗ (opens in a new tab) describes uploaded-file analysis and code-backed calculations. When calculation tools are used, inspect the method and output. Ask for the exact input rows, transformations, and formula behind a result, then reproduce important totals independently.

Use a baseline that is easy to beat honestly, such as repeating the previous comparable period. Withhold later observations, forecast them using only earlier data, and compare forecast errors with that baseline. Use several historical starting points when the available history allows it. A strong fit to the training history alone is insufficient.

Keep uncertainty labels precise. A scenario range describes the assumptions you chose; it is not automatically a prediction interval. NIST’s explanation of prediction uncertainty↗ (opens in a new tab) distinguishes uncertainty about an average from uncertainty about a new observation. Ask a qualified analyst to check statistical intervals you plan to rely on.

• Reconcile input counts and totals with the dated source file.
• Check that the forecast and actual values use the same units and periods.
• Investigate whether errors repeatedly lean in one direction.
• Report the limitations that could change the staffing or spending decision.

What should happen when actual results arrive?

Preserve the original forecast before adding actual results. Otherwise a revised sheet can erase what you believed at the time. Compare the same outcome over the same period, then distinguish an arithmetic error from an assumption that proved wrong or an event that was outside the model.

For example, lower bookings might reflect fewer available appointments rather than lower customer interest. Check the explanation with the people who operate the business. Write a follow-up task for unresolved evidence instead of declaring a cause because it sounds reasonable.

How to do this with fenbs?

1. Create a free account and add a project for the forecast you will actually review.
2. Record the approved outcome definition, scenario assumptions, and review conditions in Decisions and rules. Leave unresolved choices open for a person.
3. Add tasks for checking source totals, reproducing calculations, and comparing actual results. Keep the working model in your spreadsheet or analysis tool.
4. Connect your supported assistant using Integrations. Ask it to read the project context first and save checked findings and the next step before stopping.

fenbs does not calculate forecasts or validate statistical models. It is not the right fit when your only need is a forecasting engine; it is useful when assumptions, approvals, and follow-up work must survive the chat. People choose the plan, and assistants record what those people decided.

If you are considering another assistant for the same work, Claude vs ChatGPT and Gemini vs ChatGPT cover the separate choice of app. Keep the forecast’s evidence and checks consistent whichever one you use.