To build a seasonal demand forecast in ChatGPT, upload at least 24 months of monthly unit sales (36 is better) to its data analysis, ask for a seasonal index for each month plus a trend, and forecast the next 12 months. Test the result by hiding the last 12 months, forecasting them from the earlier data and measuring the average percentage error.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowChatGPT does this with its data analysis feature, which writes and runs Python code on your file, so the method is arithmetic you can inspect rather than a black box. It is not a forecasting product. It knows nothing about next spring's weather, a customer you are about to lose or a promotion you are planning, and it will not update itself next month. Data analysis is limited on the Free plan, which also caps uploads at three files a day, and included on Go at $8 a month and Plus at $20 a month.
Take an illustrative wholesale distributor of garden and outdoor products. It forecasts units of garden hose reels, its most seasonal line, so it can place supplier orders about six weeks ahead of the spring rush. It has three full years of monthly sales: 3,660 units in Year 1, 3,960 in Year 2 and 4,205 in Year 3.
If you need a revenue forecast for next quarter, including how to use Excel's own forecasting function, the tutorial on forecasting next quarter's sales from your history covers it. Seasonal units are a different problem: the shape of the year matters more than the total, and the test that tells you whether to believe the shape matters most of all.
How should you format a sales file for ChatGPT to forecast from?
A sales file that ChatGPT reads correctly has descriptive column headers in the first row, one record per row, and no empty rows or unrelated tables on one sheet; that is OpenAI's own guidance for data analysis, and most bad forecasts start as a bad file. For a seasonal forecast, five more rules matter:
- Units, not revenue. A price rise lifts revenue without anyone wanting more hose reels. Forecast units and multiply by price afterwards.
- Whole calendar months as real dates. Put the first day of each month in a date column, written year-month-day (for example 2025-03-01). Day-month text such as 01/03 is the commonest source of misread months.
- Zeros as 0. A blank can be read as missing; a month that sold nothing should say 0.
- Flag the months that lie. Add a column for days out of stock and a column marking promotions. A stock-out month understates demand; a promotion month overstates it.
- Full years. Twenty-four months is the minimum, 36 is better, and start in the same calendar month you end in so every month appears equally often.
The distributor's file had 108 rows: three product lines (hose reels, planters, barbecue covers) times 36 months, in .xlsx, which data analysis reads along with .xls and .csv. Spreadsheets are capped at roughly 50 MB, far more than a monthly summary needs. Its layout looked like this:
month line units stockout_days promo
2025-03-01 Hose reels 300 0 N
2025-04-01 Hose reels 520 0 N
2025-07-01 Hose reels 470 10 N
2025-03-01 Planters 610 0 Y
Customer names, invoice numbers and prices stay out of the file. A seasonal forecast only needs units by month and line, and the less you upload, the less you have to think about where it goes. If your export from the accounts or stock system arrives with merged headers, subtotal rows and stray spaces, sort that out first; the older tutorial on cleaning messy data walks through the usual fixes.
How do you check that ChatGPT read your sales file correctly?
Ask ChatGPT to describe the data before forecasting anything, then compare its totals with figures you already know. This is the cheapest check in the whole process, and it catches the errors that would otherwise poison every number that follows.
I've uploaded 36 months of unit sales for three product lines.
Before any forecasting:
1. Show the first 5 rows exactly as you read them.
2. Give the date range and the number of months for each line.
3. List any missing months, blanks, zeros or negative values.
4. Give total units per line for each 12-month year.
5. List every row with stockout_days above 0 or promo = Y.
Do not forecast anything yet.
An illustrative reply for the distributor:
The file has 108 rows: 3 lines with 36 months each, no missing months, no negative values. Hose reels totals: Year 1 3,660; Year 2 3,960; Year 3 4,205. One stock-out row: Hose reels, July Year 3, 10 days. Two promotion rows: Planters, March and April of Year 2.
The owner checked the three hose-reel totals against the stock system's annual sales report. They matched, and the rows looked right, so the forecast could start. On the first upload the rows had not looked right: the month column held day-month text such as 01/03, and ChatGPT read some of those as 3 January rather than 1 March. The yearly totals still matched, because every unit was still counted somewhere, but the missing-months list flagged gaps in March of every year and the first five rows showed spring sales sitting in January. Rewriting the dates as year-month-day and uploading again fixed it. That is why the check asks for the rows as read and for missing months, not only for totals.
How does ChatGPT build a seasonal pattern from sales history?
ChatGPT builds the seasonal pattern from each calendar month's share of total units across all the years, then turns each share into an index. A seasonal index says how a month compares with an average month: 2.00 means twice the monthly average, 0.50 means half. Multiply next year's expected total by each month's share of the year and you have a monthly forecast. The method is deliberately simple, so you can check every step in a spreadsheet and see exactly why a number came out the way it did.
Use this method for Hose reels, and show your working and the code:
1. Seasonal share for each calendar month = units in that month across
all years / total units across all years.
2. Seasonal index = share x 12.
3. Yearly growth = each full year's total / the previous year's total.
4. For July of Year 3, use 550 instead of 470 (10-day stock-out; we
estimate about 80 units of lost sales). Say where you made the change.
Show a table: month, units across all years, share, index.
The table that came back, with the July adjustment applied (illustrative figures):
| Month | Units, Years 1-3 | Share of year | Seasonal index |
|---|---|---|---|
| Jan | 400 | 3.4% | 0.40 |
| Feb | 470 | 3.9% | 0.47 |
| Mar | 950 | 8.0% | 0.96 |
| Apr | 1,600 | 13.4% | 1.61 |
| May | 2,150 | 18.1% | 2.17 |
| Jun | 1,920 | 16.1% | 1.94 |
| Jul | 1,480 | 12.4% | 1.49 |
| Aug | 1,090 | 9.2% | 1.10 |
| Sep | 700 | 5.9% | 0.71 |
| Oct | 460 | 3.9% | 0.46 |
| Nov | 355 | 3.0% | 0.36 |
| Dec | 330 | 2.8% | 0.33 |
Read it as a buyer would: May sells about 2.2 times an average month and more than six times December. Nearly half the year's hose reels (47.6%) go out between April and June. Growth was 8.2% from Year 1 to Year 2 and 8.2% again from Year 2 to the adjusted Year 3.
The July adjustment is the kind of judgement the model cannot make for you. Without it, July's share falls, next July's forecast falls with it, and the distributor orders too little and runs out again: a stock-out that teaches the forecast to cause the next stock-out. The same logic runs the other way for promotions. The distributor's planters sold heavily in a two-for-one March in Year 2; for that line the owner told ChatGPT to replace that March with the average of the other two Marches before calculating shares.
Ask to see the code once, even if you never read Python. Look for two things: the lines where the July figure is changed, and the line that decides which years feed the calculation. If either is not where you expect, ask ChatGPT to explain or adjust that step; OpenAI's guidance is to review the generated code, outputs and assumptions before relying on results.
How do you back-test a ChatGPT forecast before trusting it?
To back-test a forecast, hide the most recent full year, build the pattern and growth from the earlier years, forecast the hidden year and compare the result with what really happened. A back-test answers one question: if you had used this method a year ago, how wrong would it have been? You also compare against the laziest sensible forecast, "same month last year", because a method that cannot beat that is not worth the effort.
Now test the method honestly. Pretend Year 3 has not happened.
1. Using ONLY Year 1 and Year 2, build the seasonal shares and the growth
rate exactly as before, and forecast every month of Year 3.
2. Compare with the real Year 3 (use the real July figure, 470).
3. Show a table: month, forecast, actual, error (forecast minus actual),
absolute percentage error.
4. Give the mean absolute percentage error (MAPE), the mean absolute error
in units, and the total-year error.
5. Do the same for a naive forecast: each Year 3 month = same month of Year 2.
The illustrative result:
| Month | Forecast | Actual | Error | Abs. % error | Naive abs. % error |
|---|---|---|---|---|---|
| Jan | 141 | 150 | -9 | 6.0% | 13.3% |
| Feb | 174 | 160 | +14 | 8.8% | 0.0% |
| Mar | 365 | 300 | +65 | 21.7% | 13.3% |
| Apr | 607 | 520 | +87 | 16.7% | 7.7% |
| May | 753 | 810 | -57 | 7.0% | 13.6% |
| Jun | 675 | 720 | -45 | 6.2% | 13.9% |
| Jul | 523 | 470 | +53 | 11.3% | 2.1% |
| Aug | 388 | 400 | -12 | 3.0% | 10.0% |
| Sep | 247 | 260 | -13 | 5.0% | 11.5% |
| Oct | 163 | 170 | -7 | 4.1% | 11.8% |
| Nov | 129 | 125 | +4 | 3.2% | 4.0% |
| Dec | 118 | 120 | -2 | 1.7% | 8.3% |
Summary figures: the seasonal method's MAPE (the average of the absolute percentage errors, so a typical month's miss) was 7.9%, against 9.1% for the same-month-last-year forecast. In units, a typical month missed by about 31 against about 35. Over the whole year the method forecast 4,283 units against 4,205 actual, 1.9% high; the naive forecast came to 3,960, 5.8% low, because it ignores growth.
Check one row by hand before you accept the table. March: forecast 365, actual 300, error +65, and 65 divided by 300 is 21.7%. If a hand check disagrees with ChatGPT's figure, ask it to show that row's calculation; one wrong row usually means a wrong step for all of them.
What the back-test says about ordering hose reels
The headline MAPE hides the story that matters for buying. Look at the pattern of the errors, not only their size.
- Spring was a timing miss. The forecast was too high for March and April (misses of 22% and 17% of actual sales) and too low for May (7%). Year 3 had a late, cold spring that pushed demand into May. Across March to May together the method was 95 units high (1,725 forecast against 1,630 actual, 5.8%). No sales-history method can see weather coming, so the fix is in how you order, not in the maths.
- July's error is mostly the stock-out. The forecast said 523; recorded sales were 470 during ten days without stock. True demand was probably closer to the forecast than to the sales figure.
- Summer was nearly exact. June to August forecast 1,586 against 1,590 actual.
- The naive forecast lost on growth, not on shape. It got February exactly right and was closer in March and April, but it under-forecast eight months out of twelve because the business grows about 8% a year.
The conclusion for this line: the method beats the benchmark, but only by about a point of MAPE, and its biggest misses are early-spring timing. That points to splitting spring purchasing into two deliveries rather than chasing a cleverer model. If a back-test on your own data shows the method losing to same-month-last-year, use the naive forecast plus a growth adjustment and stop there. More reasons forecasts fail, and how to spot each one, are in why AI stock forecasts go wrong.
How do you turn the seasonal pattern into next year's forecast?
Next year's forecast runs on all three years of history with the July adjustment, plus a growth rate you have to decide for yourself. History says 8.2% a year, but the distributor had just lost a retail account that took about 3% of its hose reels, so the owner set growth at 5%. That one judgement moves the year from 4,637 units at historical growth to 4,501.
Forecast Hose reels for next year by month using all three years
(July Year 3 = 550) and growth of 5% on Year 3's adjusted total.
Show: month, forecast units, and season totals for Mar-May, Jun-Aug,
Sep-Feb. Then repeat the table at 8.2% growth for comparison.
| Season | Months (forecast units) | Season total | Back-test lesson |
|---|---|---|---|
| Spring | Mar 359, Apr 605, May 813 | 1,777 | Months can miss by up to 22%; the season missed by 6% |
| Summer | Jun 726, Jul 559, Aug 412 | 1,697 | Within 1% as a season in the test |
| Autumn and winter | Sep 265, Oct 174, Nov 134, Dec 125, Jan 151, Feb 178 | 1,027 | Monthly misses of 2% to 9% |
Turning that into a buying plan is where the back-test earns its keep. With a six-week supplier lead time, the distributor would normally order the whole spring at once. Instead it orders March and April's forecast (964 units) plus 15%, about 1,110 units, for the first delivery, and holds the May order of roughly 813 until it has seen the first two weeks of April. If spring arrives late again, the second order shrinks or slides; if it arrives early, the extra 15% covers the gap while the second order is brought forward. Summer, where the test was close, goes on a single order.
That range is a rough planning band built from one year's errors, not a statistical confidence interval. It is still far more useful than a single number presented as if it were certain.
How do you cross-check ChatGPT's arithmetic in Excel?
Two ten-minute checks in Excel catch most arithmetic surprises in a ChatGPT forecast. First, rebuild one share and the MAPE yourself. With units in column C and month numbers in column B of a sheet holding the history, the May share is =SUMIFS(C:C,B:B,5)/SUM(C:C). For the back-test, put forecast and actual side by side, add a helper column with =ABS(D2-C2)/C2 and average it; the result should match ChatGPT's 7.9% to the decimal.
Second, get an independent opinion from Excel's own forecasting function. FORECAST.ETS(target_date, values, timeline, [seasonality]) uses the AAA version of exponential smoothing, a different method from the seasonal shares above. Setting seasonality to 12 tells it the pattern repeats every 12 months; the default of 1 asks Excel to detect seasonality itself. The timeline needs a consistent step (one row per month), and the function copes with up to 30% missing points. It is available in Excel 2016 and later, including Microsoft 365. If Excel and ChatGPT disagree on a month by more than the back-test error for that season, look at that month's history before choosing either number.
If the spreadsheet side is the slow part, try The Complete Excel Course with AI Integration on Coding Liquids.
When should you not use a simple seasonal forecast?
A simple seasonal forecast is the wrong tool for product lines whose yearly shape does not repeat, because the method assumes this year's shape will look like the last few. Plenty of product lines break that assumption.
- Lumpy trade orders. The distributor's patio heaters looked seasonal until the back-test came in at a MAPE of 38%, with the naive forecast at 41%: both useless. Two orders of about 300 units from one retail chain had landed in October one year and January the next. The fix was to take that chain's orders out, forecast the rest with the seasonal method, and ask the chain's buyer for their plan directly. A conversation beat any model.
- Short or broken history. New products, discontinued lines and anything with fewer than 24 months have no pattern to find. Use a similar product's shape as a stand-in, and say so in the forecast.
- Price changes and promotions you plan. The method projects past behaviour. If next spring includes a price rise or a promotion, adjust the affected months by hand and write down why.
- Outside drivers. The Python environment behind data analysis cannot make web requests, so it cannot fetch weather, market or events data. If a driver matters, add it to your file as a column and ask ChatGPT to test whether it explains the misses.
- Hundreds of products. A seasonal forecast per line in a chat works for a handful of lines. OpenAI warns that large, complex files may not be fully analysed, and checking hundreds of back-tests by hand is not realistic. That is inventory software territory, compared in the tutorial on how AI inventory forecasting works for a small business.
A monthly routine that keeps the forecast honest
A forecast you never compare with reality slowly turns into wishful thinking. Ten minutes a month keeps it honest: add the new month's actuals to the file, log the error, and rerun the full method each quarter or whenever the running error drifts. Keep the four prompts above in a document (or in a ChatGPT project with the file) so each run uses identical instructions; if the wording changes from run to run, so can the method.
The log itself is a small table. Here are the first three months of the illustrative Year 4:
| Month | Forecast | Actual | Error | Abs. % error | Running bias |
|---|---|---|---|---|---|
| Jan | 151 | 158 | -7 | 4.4% | -7 |
| Feb | 178 | 171 | +7 | 4.1% | 0 |
| Mar | 359 | 402 | -43 | 10.7% | -43 |
March sold 43 units more than forecast (402 against 359), the opposite of the back-test year: an early spring. That is exactly the case the two-delivery plan was built for. The distributor brought the May order forward by two weeks rather than waiting for April's figures, and noted the early spring against March in the file so the next pattern calculation knows why that month ran hot.
Two signals mean it is time to rebuild rather than adjust. If the running bias keeps growing in one direction for three months or more, the growth figure is probably wrong. If one month's error is bigger than anything in the back-test, find the cause (a lost customer, a new channel, a stock problem) before the next order goes in. Neither needs a new tool; both need the owner to look at the log every month.
If you want help fitting a forecast routine into how your business actually runs, I offer one-to-one AI consultations for small businesses.
ChatGPT seasonal demand forecasting: questions owners ask
How many months of sales history does ChatGPT need for a seasonal forecast?
Twenty-four months is the practical minimum, because the method needs to see each calendar month at least twice before it can tell a seasonal pattern from a one-off. Thirty-six months is better, and it lets you back-test on a full year while still building the pattern from two. With only twelve months, use last year's same month as your forecast and adjust it by judgement.
How do I test a ChatGPT sales forecast?
Back-test it: hide the most recent full year, build the seasonal pattern from the earlier years, forecast the hidden year and compare it with what really happened using the mean absolute percentage error (MAPE). In the worked example the seasonal method scored 7.9% against 9.1% for same-month-last-year. If a method cannot beat that benchmark, do not rely on it.
Should I forecast units or revenue?
Units. Price rises, discounts and product mix changes all move revenue without any change in how many items customers want, so a revenue pattern blends demand with pricing decisions. Forecast units per product line, then multiply by the prices you plan to charge if you also need a revenue figure for a cash flow forecast or a budget.
Can ChatGPT keep the forecast updated automatically each month?
Not in this setup. Each run is a fresh analysis of the file you upload, so next month you add the new month's actuals, upload the updated file and run the same prompts again. Keeping the prompts and the method notes in one place makes that a ten-minute job. If you need forecasts that refresh themselves across hundreds of products, that is a job for inventory software.
Is my sales data used to train OpenAI's models?
On ChatGPT Business, Enterprise and Edu, OpenAI says it does not use workspace content for training by default. On Free, Go, Plus and Pro, training on your conversations depends on the setting called Improve the model for everyone, under Settings and then Data controls. Check it before uploading customer-level sales data, or upload only monthly totals by product line.
Further reads
- How Small Shops Can Use AI to Forecast Stock and Reorder — Turn a forecast into reorder points for a small shop.
- Is Cin7 ForesightAI Forecasting Worth It for a Small Maker? — When paid forecasting inside inventory software is worth it.
- Warehouse Labour Planning: AI Forecasts Shifts From Order Volumes — Turn forecast volumes into warehouse shift plans.
- ChatGPT Excel Dashboard: A KPI Page Built From Last Month's Sales — Put forecast against actual on a one-page Excel dashboard.
- How a Cloud Kitchen Can Forecast Demand by Delivery App With AI — How to forecast orders per delivery app and daypart, measure promo and weather effects, and turn the result into a prep sheet your kitchen can follow.
- Hardware Store Marketing With AI: Trade Accounts and DIY Weekends — Write a trade-account pitch, plan DIY demo weekends with project sheets, and time seasonal stock posts from last year's sales, using AI you check yourself.
Show 8 more related tutorials
- Hotel Breakfast Food Waste: Forecast Covers From Occupancy — A worked breakfast-covers forecast from occupancy, with an accuracy test, buffet replenishment rules and a five-minute waste log for a small hotel.
- How AI Can Help in Sales: A Gift Shop's Bundle and Follow-Up Plan — A gift shop's AI sales plan built on its own till data: three priced bundles, gift-note wording and four follow-up messages for one holiday month.
- How to Forecast Ice Cream Tubs by Flavour, Weather and Weekday — Build a simple sheet that turns your sales log and the weather forecast into tubs per flavour, with a reorder rule and a spoilage check.
- Can AI Forecast Holiday Demand and Staffing for Pet Boarding? — Turn three years of boarding bookings into a holiday forecast, a staffing sum and waitlist rules, with the checks that keep AI honest.
- Rental Car Companies Using AI: Seven Uses a Small Fleet Can Copy — Seven AI jobs a 20-car rental fleet can copy, each with a worked example, a cost and a check, from rate rules to after-hours calls.
- AI Staffing Tools for Workforce Planning: Forecast Client Demand — Turn two years of placement history into a seasonal demand forecast and a bench target, using a spreadsheet and AI for the clean-up and cross-check.
- AI Demand Planning Software: Built-In Add-Ons vs a ChatGPT Spreadsheet — Katana and Cin7 add-ons versus a DIY forecast: a backtest on last year's sales shows which one a small product business needs.
- ChatGPT for Excel: Clean and Summarise a Monthly Sales Export — Clean a monthly transaction export with a row-level exception log and a sales summary that reconciles to the untouched source.
Browse all AI-for-business tutorials
Sources: OpenAI help pages on data analysis, file uploads and data controls; the ChatGPT pricing page; Microsoft's FORECAST.ETS and forecasting functions reference, all checked 17 November 2026.
