Gains:
- Ability to position AI as an assistant that generates hypothetical scenarios in income/expense forecasting
- Ability to write forecast prompts with range and sensitivity based on historical data and explicit assumptions
- Ability to recognize the limitations of AI predictions and blend them with human judgment
"What will be our turnover next year?" is one of the most difficult questions for the finance team. Forecasting is making a prediction about the future based on past data and assumptions. Artificial intelligence (AI) can be a powerful assistant here; But it is necessary to accept a fact from the very beginning: the model does not know the future. It only processes patterns and assumptions you give. In this unit, we will learn to position AI in its correct role in forecasting and recognize its limits.
What Does AI Do and Doesn't Do in Forecast?
It does: Summarizes the past trend, recognizes seasonality (repeated fluctuation in certain months), builds scenarios with different assumptions, explains the forecast logic, and puts the forecast into a narrative.
It does not: It cannot give an accurate future, it cannot predict external shocks (exchange rate crisis, regulation change), it cannot know a fact that is not in the data. The sentence "you will grow 18% next year" given by a chatbot is not a prophecy, but a mathematical reflection of the assumptions you make.
Tip: Never ask the model for a "single number." Always ask for a range and what assumptions that range is based on. The sentence "14-15M in the base case, condition: current growth rate continues" is much more honest and useful than the single "14.6M".
Step by Step: Building a Solid Forecast
- Give historical data. Occurring for at least a few periods; monthly if possible.
- Write the assumptions clearly. Growth rate, seasonality, known changes.
- Specify method. Simple trend, seasonally adjusted or scenario based?
- Ask for range and scenario. Optimist / base / pessimist.
- Ask for sensitivity. “Which assumption changes the outcome the most?”
Weak Prompt / Strong Prompt
Weak prompt: Predict my turnover next year. This year we made 12M.
This produces a single number with no basis; It is unclear what assumption it is based on, which is dangerous for the decision.
Strong prompt:Your role: a financial planning analyst.Task: Forecast the next 12 months from the following monthly income data.Method:1) First summarize the trend and seasonality in historical data.2) Set up 3 scenarios with the following assumptions: - Base: average growth rate of the last 12 months continues - Optimistic: growth +3 points - Pessimistic: growth -5 points3) Give 12-month total and monthly range for each scenario.4) "Sensitivity" section: which assumption changes, the result changes the most?Constraint: List the assumptions explicitly. State that the estimate is a prediction and not a certainty. Note that you cannot model external shocks.<data>January..December monthly turnover figures</data>
This prompt; Instead of an exact number, it produces a reasoned range, scenarios and a sensitivity analysis such as "most vulnerable assumption collection/growth rate". The decision maker makes decisions by seeing uncertainty.
Seasonality is the most frequently overlooked topic in forecasting. Sales of many businesses rise and fall regularly in certain months: an ice cream maker peaks in the summer, a souvenir shop peaks around the holidays. A flat trend line cannot see this fluctuation and will set up the stock, personnel and cash plan incorrectly. If you give the AI enough historical data (ideally at least 24 months), it will notice this repeating pattern and adjust the forecast accordingly.
Prediction Methods and the Role of AI
Method
When is it appropriate?
AI contribution
border
simple trend
Stable, unseasonal data
Gets the trend fast
Can't see the breaks
seasonal correction
Sales fluctuating by month
Explains the seasonal pattern
Too attached to the past
Scenario based
A period of high uncertainty
Tests hypotheses quickly
Depends on the quality of the assumption
Driver-based
If income depends on several factors
Establishes factor relationship
Human verifies the relationship
Driver-Based Prediction
Predicting turnover with a single growth percentage is often crude. The more robust way is to divide revenue by its drivers: revenue = number of customers × average sales per customer × repeat frequency. Predicting and multiplying each driver separately gives a prediction that is both more accurate and more questionable; because you clearly see which assumption has changed.
Estimate my income based on drivers. Revenue = number of active customers × average monthly revenue per customer (ARPU). Extract these two drivers separately for the past 12 months, interpret the trend of each, then forecast the next 12 months by projecting the two drivers separately. Write the assumption clearly for each driver; Specify which driver is more unclear. Use only numbers in the data, show formulas.
The power of this method is that instead of saying "turnover increases by 15%", you say "number of customers increases by 8%, revenue per customer increases by 6%". The two can be questioned separately; If one is unrealistic, you adjust the estimate from there. Management also asks "what assumption do we rely on?" can answer the question concretely.
Sensitivity and Assumption Discipline
The quality of the forecast is the quality of the assumptions. Always keep the assumptions visible to the AI and go back and update them as they happen (rolling forecast, that is, a rolling forecast that is updated every month).
List ALL the assumptions you used in the forecast below in a tabular form: assumption | value | source/reason | If this value changes by 10%, its effect on the annual result. Tick the 3 most fragile assumptions.
When presenting the AI's predictive output to management, include an "assumption card": the three most critical assumptions, their values, and their sources. Thus, the prediction is no longer a black box; Everyone sees what they are betting on. And when the realization deviates, you can quickly diagnose which assumption didn't hold, which improves the next forecast.
Caution: Do not defend the AI-generated prediction by saying "that's what the model said." The responsibility for the prediction lies with the person presenting it. Always present to management with the framework of “this is a scenario based on these assumptions, the actualization may differ”.
Mini Cases
Case 1 — Forecast that sees seasonality. A retail company fed 24 months of sales data to AI. The model noticed a steady 40% jump in November-December and adjusted its forecast for next year accordingly. If a flat trend had been used, the year-end stock plan would have been seriously wrong.
Case 2 — The trap of the odd number. A startup bases its budget on AI's "18M next year" forecast. But this was based on the assumption that "growth continues as it is". 11M was realized when two major customers left. Once the lesson was learned, the team moved on to scenario range; The next year's forecast band covered what actually happened.
Case 3 — Sensitivity changed the decision. “The most vulnerable assumption is the raw material price,” AI said in a manufacturer’s forecast, showing that a 10% price increase would reduce annual profits by 22%. Management then signed a fixed-price annual contract with the supplier. The real value of the forecast was not the number, but this sentiment insight.
Case 4 — Rolling forecast wins. One retailer updated its annual forecast with actuals each month, rather than setting it once and forgetting it. In the third month, it was seen that the "growth 12%" assumption was actually 7%; The forecast was corrected early and the year-end inventory and staffing plan was adjusted accordingly. If the fixed forecast had been maintained, there would have been a huge overstock and cash crunch at the end of the year. Lesson: a forecast is a living document, it is not set up once and put away.
Common mistakes
- Asking for an odd number. Forecasting without range and script creates false certainty.
- Hiding assumptions. Invisible assumption means a prediction that cannot be questioned and updated.
- Relying too much on the past. The pattern extends the past pattern; cannot see fractures and external shocks.
- Mistaking a prediction as a prophecy. AI does not know the future; It just takes your assumptions into account.
- Not making a rolling forecast. Set the forecast once and forget it; not updating as it happens.
In summary
- It is a powerful assistant in AI forecast: extracts trend and seasonality, builds scenario, explains logic; but it does not know the future and cannot see external shocks.
- Always ask for a hypothetical range and optimistic/baseline/pessimistic scenarios rather than a single number.
- Have the assumptions clearly listed and identify the most vulnerable assumption through sensitivity analysis.
- The responsibility for the prediction lies with the person presenting it; Present to management with a "scenario, not a certainty" framework.
- Update the forecast as it happens with Rolling forecast; Install it once and don't forget.
Application task
Obtain at least 12 months of actual revenue data (actual or sample). Create a three-scenario 12-month forecast with a powerful prompt template. Then remove the three most fragile assumptions with the sensitivity prompt. Manually verify the monthly growth rate of the model's base case once and summarize the forecast on a slide with a "scenario, not certainty" framework.
checklist
- [ ] I have provided sufficient historical data (preferably monthly).
- [ ] I listed the assumptions explicitly and had them listed in the model.
- [ ] I wanted a scripted range, not a single number.
- [ ] I identified the most fragile assumption through sensitivity analysis.
- [ ] I manually verified the growth calculation of the base case.
- [ ] I presented the estimate in a "scenario, not certainty" framework.