Workflow

Anyone can draw a line. The question is how wrong it is

The M-competitions established something the forecasting industry still finds uncomfortable: simple methods, properly evaluated, routinely beat complicated ones. SheetAI builds your forecast in the sheet, tests it against a naive baseline on data it did not see, and reports the error, so what you present is a number with a stated confidence rather than a curve with a nice slope.

Try it on your sheet See real workflows

Audit-ready trailNo data trainingReversible per action

42,840

Walmart series in the M5 competition

The M5 competition, run by Makridakis, Spiliotis and Assimakopoulos, asked forecasters to predict 42,840 series of hierarchical unit sales from Walmart.

Source: Makridakis et al.: M5 accuracy competition, International Journal of Forecasting (2022)

45%

Confident in their sales forecasts

Gartner's State of Sales Operations Survey found that only 45% of sales leaders and sellers have high confidence in their organisation's forecasting accuracy.

Source: Gartner: State of Sales Operations Survey (2020)

60,000+

Forecasts checked for human adjustments

Fildes, Goodwin, Lawrence and Nikolopoulos examined over 60,000 forecasts from four supply-chain companies. Small manual adjustments often reduced accuracy, and upward adjustments leaned optimistic.

Source: Fildes et al.: Effective forecasting and judgmental adjustments, International Journal of Forecasting (2009)

Forecasting & Projections: illustration for SheetAI's forecasting & projections use case

In the M4 competition, statistically simple methods matched or beat most complex ones.

Across 100,000 series, Makridakis and colleagues found that sophistication did not reliably buy accuracy, and that combinations of simple methods performed exceptionally well. The lesson is not that models do not matter. It is that evaluation matters more, and most business forecasts have never been evaluated at all.

Source: Makridakis et al.: The M4 Competition

This page is about projecting a number forward defensibly: demand by SKU, revenue by segment, cash position by week, headcount cost by quarter, ticket volume by month. SheetAI builds the forecast inside your workbook, and (this is the part that distinguishes it from drawing a trendline) holds back the most recent periods, forecasts them, and compares the result against both the actuals and a naive baseline. You get the projection, the error metric, and an honest statement of whether the model beat simply repeating last year. It will not tell you the future, it will not price a security, and it will not manufacture confidence intervals for a series that has no stable structure.

What the forecasting research actually says

Four findings from the literature that contradict how most business forecasts are built.

Complex methods reliably beating simple ones in M4

Across 100,000 series, pure machine-learning approaches largely underperformed statistical benchmarks; the winning entries were hybrids and combinations. Sophistication is not a proxy for accuracy.

Makridakis et al., M4 (2020)

Value of combining several simple forecasts

High

Combination has been one of the most robust findings in the field for fifty years: averaging several reasonable methods usually beats picking one, and it reduces the variance of being wrong.

M-competition literature

Business forecasts ever tested against a naive baseline

Few

The comparison costs one column and settles whether the model adds anything. It is skipped almost universally, which is why so many forecasting processes have unknown value.

Practitioner experience

Forecast accuracy reported with an error metric

Rare

A projection presented without MAPE, MAE, or an interval is a point estimate wearing a suit. The uncertainty exists whether or not it is on the slide.

IBF / forecasting practice surveys

The pattern: the discipline of forecasting is mostly the discipline of evaluation. Building a projection takes minutes; establishing whether it is better than the obvious alternative takes one more step that almost nobody performs, and it is the step that determines whether the number should influence a decision at all.

Anatomy of a forecast done properly

The sequence that separates a defensible projection from a line with a slope.

Phase 1: Look at the series before modelling it

›Plot the history; check for level shifts and structural breaks
›Identify seasonality and its period
›Find the outliers and decide what they were

⚠ A promotion, a stockout, or a pandemic quarter left in the history as if it were normal demand will bend every projection built on it.

Phase 2: Establish the baseline

›Naive: next period equals last period
›Seasonal naive: equals the same period last year
›Compute the baseline's error on the holdout

⚠ Skipped almost always, which means nobody ever learns that the elaborate model is losing to last year's number repeated.

Phase 3: Fit candidate methods

›Trend and seasonal decomposition
›Exponential smoothing variants
›Regression on the drivers you actually have

⚠ Method selection absorbs the attention because it feels like the real work. It is the part the evidence says matters least.

Phase 4: Backtest on held-out periods

›Hold back the most recent periods
›Forecast them without having seen them
›Compare each method against the baseline

⚠ In-sample fit is seductive and meaningless. A model that fits history perfectly and fails on the holdout has memorised, not learned.

Phase 5: Present with the uncertainty attached

›Report the error metric alongside the number
›State the assumptions and what would break them
›Track actuals against forecast next period

⚠ The single point estimate is what gets remembered, and the caveat is what gets cut for space, which is how a forecast becomes a promise.

Cautionary tale

A forecast without an interval is a promise nobody meant to make

In 2008, Fannie Mae disclosed a $1.136 billion error in its reported shareholder equity, caused by what the company described as an honest mistake in a spreadsheet related to the implementation of a new accounting standard. No model was too simple or too complex; a formula produced a number, the number was published, and it was wrong by more than a billion dollars. The point is not the arithmetic. It is that a single figure carried into a report acquires an authority entirely disconnected from how it was produced, and once it is in the deck nobody re-derives it.

The lesson: The same dynamic applies to any projection. A forecast becomes a commitment the moment it appears on a slide without an error band, because the range is what tells a reader how much weight to place on it. SheetAI reports the backtest error alongside the projection, states the baseline it beat or failed to beat, and writes the assumptions into the sheet, so the person reading the number can see what it is worth.

Source: Reuters: Fannie Mae reports $1.1 billion spreadsheet error

Who it's for

If you are asked for a number about next quarter, this is the page.

FP&A Analysts

Rebuild the revenue and cost projection every month with no measurement of whether last month's was any good.

Demand Planners

Forecast hundreds of SKUs where a handful of items carry most of the value and the rest are noise.

Founders & Finance Leads

Need a cash runway projection that survives a board question about its assumptions.

Operations Managers

Staff to expected volume and pay for both over- and under-forecasting.

Sales Operations

Roll up a pipeline forecast whose historical accuracy nobody has ever measured.

Real forecasting workflows

The prompt, the method, and the honest accuracy statement that comes with it.

forecasting_workbook.xlsx - SheetAI

You ask

“Analyse the monthly series in Revenue!A:B covering 2019 to now. Identify the trend, any seasonality and its period, structural breaks, and outliers. Tell me which periods you would exclude or adjust before forecasting, and why. Do not build a forecast yet.”

SheetAI does

✓Decomposes the series into trend, seasonal, and residual components.
✓Tests for level shifts rather than treating the history as homogeneous.
✓Identifies outliers and characterises them rather than deleting them.
✓Recommends specific adjustments with reasoning, before any model is fitted.

Formula written

=AVERAGE(OFFSET($B2,-11,0,12,1))

Result

Clear annual seasonality with a December peak, a permanent level shift in April 2023 coinciding with a pricing change, and three outlier months identified as one-off contracts, all of which would have quietly distorted a naive trend fit.

Everything you need to forecast honestly

Decomposition, baselines, backtesting on held-out periods, and error metrics reported next to the number. Python runs the statistics; your sheet holds the result.

✓Trend, seasonality, and break detection

✓Naive and seasonal-naive baselines by default

✓Holdout backtesting with MAPE and MAE

✓Scenario ranges rather than point estimates

✓ABC segmentation for large catalogues

✓Bias and accuracy tracking across periods

Plays well with your stack

ERP and accounting exports (NetSuite, Xero, QuickBooks, Sage)
CRM pipeline exports (Salesforce, HubSpot)
Ecommerce order history (Shopify, WooCommerce, Amazon)
Inventory and WMS extracts
Bank transaction downloads
Any monthly or weekly series in a sheet

What a monthly forecast cycle looks like

A representative planning team producing a rolling demand and revenue forecast, before and after the modelling and backtesting move into the sheet.

Before SheetAI

~28 hours

Mon
Pull the history; clean it; argue about which periods are anomalous
~6h
Tue
Rebuild the model in a copy of last month's workbook
~7h
Wed
Adjust for known events by hand; document nothing
~6h
Thu
Build the deck; present a single number with no error band
~5h
Fri
Answer challenges by re-deriving figures from memory
~4h

✓

With SheetAI

~7 hours

Mon
Series analysed; breaks and outliers surfaced with reasons
~1h
Tue
Baselines and candidate methods backtested on a holdout
~2h
Wed
Scenarios built; assumptions written into their own sheet
~2h
Thu
Present with MAPE attached; last month's accuracy reviewed
~2h

~75% reduction on a representative monthly cycle, and a backtest that previously never happened at all.

What SheetAI will not do

Forecasting attracts more overclaiming than any other analytical task. These are the lines.

Predict the future

It extrapolates structure that exists in your history. A forecast is a statement of what continues if nothing changes, and things change. Every projection it produces comes with an error metric and an assumption list precisely because the number alone would imply a confidence that no method can support.

Give financial or investment advice

It will not forecast a share price, value a security, or tell you whether to make an investment. Business planning forecasts (demand, revenue, cash, headcount) are what this is for. Anything that constitutes a regulated financial recommendation belongs with a person who is authorised to give one.

Forecast a series that has no structure

Intermittent demand, brand new products with three data points, and series dominated by one-off events cannot be forecast meaningfully by any method. It says so and recommends a different approach (reorder points, analogue-based planning) rather than producing a confident number that is really a random walk.

Hide the baseline comparison

Every forecast is reported against a naive benchmark, including when the naive benchmark wins. A model that loses to repeating last year's figure is useful information and it will be stated plainly, because the alternative is a process that consumes a week a month and nobody knows whether it helps.

“

The uncomfortable moment was the first backtest. Our model was losing to seasonal naive on about a third of the categories, a model we had been running for four years. We fixed two of those categories and stopped modelling the rest. The forecast got better and the process got shorter, which I did not think were compatible.

Demand Planning Manager, consumer products (340 SKUs)

Frequently asked

What people ask before they put a projection in front of a board.

What forecasting method does it use?

›

Whichever suits the series, and it tests rather than assumes: decomposition with trend and seasonality, exponential smoothing variants, regression against drivers you supply, and combinations, which the M-competition literature has shown repeatedly to be one of the most robust approaches available. The method matters less than the evaluation, which is why every forecast is backtested against naive and seasonal-naive baselines.

How accurate will my forecast be?

›

Nobody can answer that in advance, and be sceptical of anything that claims to. What you can have is a measured answer: hold out your most recent periods, forecast them blind, and compute the error. That gives you a MAPE for your data with your seasonality and your volatility. A stable monthly revenue series might come in under 10%; intermittent SKU-level demand might be 40% and that is not a failure, it is the nature of the series.

Why does it keep comparing against a naive forecast?

›

Because it is the only honest test of whether the model adds value. A naive forecast is free: next period equals this period, or the same period last year. If an elaborate model cannot beat that, the elaborate model is worse than doing nothing, and you have just found out for the cost of one column. It is remarkable how often this comparison has never been run.

Can it forecast with a short history?

›

It can produce something, and it will tell you how much to trust it. Detecting annual seasonality genuinely needs two or more full cycles; with eighteen months you are estimating a seasonal pattern from a single observation of each period. It states these limitations rather than producing a smooth curve that implies more information than exists.

Does it handle promotions and one-off events?

›

If you tell it about them, yes; mark the affected periods and it can model them as interventions or exclude them from the baseline fit. What it cannot do is know that March 2024 was distorted by a promotion unless something in the data says so. Unmarked one-offs left in the history are the most common cause of a quietly wrong forecast.

What about external drivers like price or weather?

›

Supply them as columns and it can build a regression against them, reporting how much of the variation they actually explain. Two cautions: a driver that explains history is not necessarily a driver that predicts, and you need a forecast of the driver itself to forecast with it, which is often a harder problem than the one you started with.

Can it produce confidence intervals?

›

Yes, and it prefers scenario ranges derived from backtest error over intervals derived purely from model assumptions, because the second kind is reliably too narrow. Either way the range should be presented alongside the point estimate rather than dropped for space; the range is what tells your reader how much weight to place on the number.

Is this a replacement for a planning system?

›

No. If you are running S&OP across thousands of SKUs with multi-echelon inventory optimisation, that is a planning platform's job. This is for the forecast that currently lives in a spreadsheet, which for most organisations is the revenue forecast, the cash projection, and the long tail of items the planning system does not cover well.

Sources & further reading

Every benchmark and statistic on this page is drawn from publicly available research. We cite our sources because we read theirs.

Get started

Stop drawing trendlines.
Start measuring error.

Drop your history into a sheet and ask for a forecast with a backtest. You will find out in minutes whether your model beats repeating last year.

All use cases

Pay as you go, no subscription required