Communicating forecasts: fan charts, scenarios and the model behind them
TL;DR
Difficulty: Advanced
A forecast is a range of possible futures, not a line.
- Draw it as a fan chart: a central forecast with a 50% and a 90% band that widen with the horizon, because errors pile up.
- The model decides the band: time series models give intervals that grow with time, regression gives intervals that grow away from the data you have seen, machine learning often gives a single number until you add quantile models, backtest errors or conformal prediction.
- Bands usually come out too narrow, so treat them as the least doubt you should have.
- Add up to three scenarios with written assumptions, put numbers next to probability words, label the band so nobody reads it as size, and show last year's forecast against what happened.
A forecast shown as one confident line invites people to trust it blindly or to ignore it. Here is how to show a forecast honestly: fan charts, the maths that makes them widen, why your model decides which band you are allowed to draw, scenarios, probability words and the track record that earns a board's trust.
The question this chapter answers
You have a forecast. Maybe you built it, maybe a tool built it, maybe someone handed it to you in a spreadsheet. Now you have to show it to people who will make decisions with it: a board, a bank, a buyer ordering stock for next summer.
Two things can go wrong. They can trust it blindly, as if next year were already written down. Or they can ignore it, because "forecasts are always wrong anyway". Both are expensive. The first orders too much stock. The second orders by gut feeling.
This chapter is about the chart that sits between those two mistakes. It builds on showing uncertainty honestly, which explains intervals and error bars for a single estimate. Here the estimate stretches into the future, and that changes the picture. How to build the forecast itself is covered in the forecasting guide and in forecasting demand with autocorrelation and seasonality. This chapter is about how to show it, and why the kind of model you used changes what you are allowed to draw.
Throughout, I use an invented business: Shannon Outfitters, an outdoor gear brand in Limerick that sells online (shannonoutfitters.example) and through one shop on O'Connell Street. Its analyst, Sam, has to present next year's sales forecast to the owner, Aoife, and the board. Every Shannon Outfitters number in this article is made up for teaching.
A forecast is a range of futures, not a line
A forecast answers the question "what will happen?" with something more honest than one number. It says which outcomes are likely, which are possible, and which would surprise you. In statistics that is called a distribution: a spread of possible values, each with its own chance.
When you squeeze a distribution into one number, usually the middle one, you throw most of the information away. The middle is the single most likely value, but on its own it is rarely what happens. The question the buyer actually has is "how bad could it get, and how good?", and the single number cannot answer it.
Example: one number for December
Sam's model says Shannon Outfitters will sell €189k in December 2026. That is the middle of the forecast, called the central forecast or the median.
Here is what the same model says in full:
- there is a 50% chance December lands between €183k and €195k,
- there is a 90% chance it lands between €175k and €203k,
- which leaves a 10% chance it lands outside that range, 5% below €175k and 5% above €203k.
Now picture the buyer ordering down jackets for December. "€189k" invites an order sized for exactly €189k. The range tells a different story: if December comes in at €175k, there will be about €14k of sales worth of jackets left over. If it comes in at €203k, the shelves will be empty in the busiest fortnight of the year. The decision is about the edges, and only the range shows the edges.
The single line: why it misleads
The most common forecast chart in any board pack is a line that carries on from the history into the future, perhaps dashed, perhaps in a different colour. It looks careful. It is the opposite.
A single line says two false things at once. It says the future is as certain as the past, because the line is drawn with the same weight on both sides of today. And it says every month of the future is equally certain, although next month is far easier to predict than next October.
Example: Sam's first draft
Here is Sam's first draft for the board. The history is real (well, invented, but treat it as real), and the forecast continues it month by month.
Aoife reads it the way anyone would: "So June 2027 is €181k." A board member writes €1,874k into next year's budget, the sum of the twelve forecast months. Nobody asks how sure Sam is, because the chart never suggested there was anything to ask.
Fan charts: how a central bank learned to draw doubt
The best known answer to this problem comes from the Bank of England. Between 1993 and early 1996 it published its inflation forecast as a central line with one shaded area around it, based on the size of past forecast errors. The Bank's own staff later wrote that the shaded area was often misread as hard upper and lower limits, and that small moves in the central line got far too much attention compared with the uncertainty around it.1
From February 1996 the Bank switched to a new picture: the fan chart.1 Instead of one line and one band, it draws the whole distribution as layers of colour. The darkest band in the middle holds the outcomes judged most likely. Each lighter pair of bands outside it adds another slice of probability, in steps of 10 percentage points, until 90% of the distribution is covered.1 The further ahead you look, the wider the bands and the paler the colour. A fan chart for growth followed in November 1997.1
The name says it: the forecast fans out from today like a hand of cards. The width is the message. A narrow fan says "I know this well". A wide fan says "plan for a range".
Example: Sam's fan chart
Sam redraws the forecast with two bands, which is plenty for a board: a 50% band (the darker one) and a 90% band (the lighter one). The central forecast stays as a line, now clearly sitting inside its uncertainty.
Read it like this. In November 2026, one month out, the 90% band runs from €155k to €175k: a range of €20k. By June 2027 it runs from €153k to €209k, and by October 2027 from €105k to €173k: almost €70k wide. Sam has not become worse at forecasting between November and October. The future simply gets harder to see the further away it is, and the chart now says so.
The board's reaction changes too. Nobody writes €181k into the June budget as a fact. Someone asks: "What would have to change if June came in at €153k?" That is the question a forecast is for.
Why the fan widens: the maths
The fan does not widen because someone drew it that way. It widens because errors pile up. A forecast one month ahead has to get one month right. A forecast twelve months ahead stacks twelve months of surprises on top of each other, and they do not cancel out neatly.
For the simplest forecasting method, the naive method, which says "next month will look like this month", the spread of the forecast error grows with the square root of the horizon.2 Most real models widen in a similar way.2
Where is the estimated standard deviation of the forecast error months ahead, is the standard deviation of the one month ahead error (estimated from the model's past errors, the residuals), and is the forecast horizon in months. A standard deviation measures how far values typically sit from their average; Foundations: spread explains it from scratch.
An interval is then the central forecast plus and minus a multiple of that spread:
Where is the central forecast for months after the last observed month , is the multiplier for the coverage you want (about for a 50% interval, for 90%, for 95%, assuming the errors follow a normal distribution), and is the spread from the formula above. Foundations: confidence intervals shows where these multipliers come from.
Example: three months, three widths
Sam's model has a one month ahead error spread of , in thousands of euro. To keep the sums visible, Sam's invented model follows the square root rule exactly.
One month ahead, November 2026, central forecast €165k:
With , the 90% interval runs from €155.1k to €174.9k.
Three months ahead, January 2027, central forecast €126k:
With , the 90% interval runs from €108.9k to €143.1k.
Twelve months ahead, October 2027, central forecast €139k:
With , the 90% interval runs from €104.8k to €173.2k. The 50% interval uses instead: , from €125.0k to €153.0k.
Notice that the horizon went up twelve times but the width went up only about three and a half times (). That is why fan charts curve outwards and then flatten a little, rather than opening like a straight sided cone.
One honest warning belongs on every fan chart: these intervals usually come out too narrow, not too wide. They capture the noise the model has seen, not the shocks it has not, such as a supplier going bust or a wet summer. The large M4 forecasting competition, which compared 61 methods on 100,000 series, found that forecasting methods tend to underestimate uncertainty considerably, and only the two most accurate methods got their 95% intervals right.34 Treat the outer band as a floor for your doubt, not a ceiling.
The model decides which band you can draw
Here is the part most forecast charts skip. Not every model produces the same kind of uncertainty, and some produce none at all. Before you draw a band, you have to know where it comes from.
Time series models: the band grows with the horizon
A time series model forecasts a number from its own past: last month, the same month last year, the trend. ARIMA (AutoRegressive Integrated Moving Average) and exponential smoothing are the two classic families, both explained in the forecasting guide. Because they describe how the series moves from one month to the next, they also know how errors pile up from one month to the next. Their prediction intervals come out of the model itself and usually widen as the horizon grows.2 This is the natural home of the fan chart.
Example: the fan Sam already drew
Sam's forecast is an exponential smoothing model with a seasonal pattern: December high, February low, a summer peak for tents and boots. The fan chart above is exactly what it produces. Sam does not need any extra machinery. The caption says: "Bands from the model's own 50% and 90% prediction intervals." That one sentence tells a careful reader where the uncertainty came from.
Regression: the band is narrowest at the centre of your data
A regression forecasts from drivers instead of from the calendar: marketing spend, price, the number of rainy weekends. Its uncertainty behaves differently. It does not care how far ahead in time you are. It cares how far your driver is from the values the model has already seen.
For a simple regression with one driver, the approximate 95% prediction interval is:5
Where is the predicted value at driver value , is the standard error of the regression (the typical size of its errors), is the number of observations the model was fitted on, is the average of the driver in that data, and is the standard deviation of the driver. The term is the important one: it grows as you move away from the average, so the interval widens.5 Foundations: regression builds the model step by step.
The shape on a chart is an hourglass lying on its side: narrow in the middle of your experience, wide at the edges.
Example: what if Aoife doubles the ad budget?
Sam has 24 months of data on monthly online marketing spend and online sales. Spend ranged from €8k to €22k, averaging with (in thousands). The fitted model is online sales spend, with .
Aoife asks what happens at €40k of spend. Step by step, for a 90% interval (multiplier ):
That is the distance term, with , , and . Next the square root and the central prediction :
Finally the interval, with :
So with the 90% interval at €40k spend runs from €143.5k to €192.5k. At €15k, the centre of the data, the same sums give . The band is more than half as wide again.
And here is the caption Sam adds, because the formula has a blind spot: "Above €22k of spend this is extrapolation. The band assumes the straight line still holds, and nobody has checked that." Ads usually get less effective as you spend more, because you run out of the most interested people first. The interval measures noise around the line. It cannot measure whether the line itself bends. A chart that draws a neat band over territory the data never visited should say so in words.
Machine learning: often a single number and nothing else
Many machine learning models, such as gradient boosted trees or neural networks, are trained to give one best guess per month. Out of the box, they produce a point forecast with no band at all. That is not a flaw in the model, it is simply what it was asked to do. But it means the chart has nothing honest to draw around the line until you add it.
There are three common ways to add it:
- Quantile models. Instead of training the model to hit the middle, you train it to hit a chosen percentile, for example the 5th and the 95th. The idea goes back to quantile regression in 1978.6 Popular libraries support it directly, and the
scikit-learndocumentation shows how to build a prediction interval from three gradient boosting models trained this way.7 - Backtest errors. You replay history: pretend it is a past month, forecast, compare with what actually happened, repeat. The spread of those errors is your band.
- Conformal prediction. A formal version of the backtest idea. It wraps any model, however complex, and produces intervals that contain the true value at the rate you choose, without assuming a particular distribution for the data.8 The guarantee assumes the future behaves like the period you calibrated on, which is worth saying out loud in a business that is changing.
The simplest version, called split conformal prediction, picks one of the sorted past errors as the width of the band:8
Where is the number of past errors you collected on data the model was not trained on, is the share of outcomes you accept falling outside the band ( for a 90% interval), means round up to the next whole number, and is the position, counting from the smallest, of the error you use as the band width. Foundations: confidence intervals covers the coverage idea behind it.
Example: putting a band on Sam's machine learning forecast
Sam also tests a gradient boosting model that uses the weather forecast and the marketing calendar. It says March 2027 will bring €145k, and nothing else.
Sam backtests it on 20 past months, each time forecasting three months ahead, and records the absolute percentage error of each. Sorted from smallest to largest, the errors run from 1.2% to 21.0%.
For a 90% band, :
With , the 19th smallest error sets the band: 15.8%. So the band is the forecast plus and minus 15.8%:
With a point forecast of , the 90% interval for March 2027 runs from €122.1k to €167.9k. The chart below shows the 20 errors, with the one that sets the band highlighted.
Two notes for the caption. This band is for three months ahead only. For six or twelve months ahead, Sam repeats the backtest at that horizon, and the band will be wider, which gives a fan again. And with only 20 errors the band itself is rough: one unusual month in the history moves it. Sam writes both into the notes under the chart rather than pretending to precision.
Combining models: why the average often wins
When Sam has three models, a time series model, a regression and a machine learning model, the tempting move is to pick the one that looks best. The evidence points the other way. In the M4 competition, 12 of the 17 most accurate methods were combinations of several methods, mostly simple statistical ones.4 Different models make different mistakes, and averaging lets some of those mistakes cancel.
Combining also gives you a picture you can show. Each model's forecast is one possible future. Drawn together as thin lines, they show how much the models agree. Where the lines bunch, the models agree. Where they spread, nobody knows, and the chart says that without any statistics at all.
Example: June 2027 from three models
For June 2027, Sam's three models say:
- exponential smoothing: €181k,
- regression on planned spend: €174k,
- gradient boosting: €186k.
The simple average is:
The combined forecast is , so Sam uses €180k as the combined central forecast and shows the three individual forecasts as small marks next to it. The spread between them, €12k, is a quick sanity check on the fan chart: if the 50% band were narrower than the disagreement between three reasonable models, it would be too confident. Here the 50% band for June is about €23k wide, so it passes.
Scenarios: best, base and worst, with the assumptions written down
A fan chart answers "how uncertain is this, given everything stays roughly as it was?". It cannot answer "what if something specific changes?". For that, you need scenarios: separate forecasts, each built on a stated assumption.
The rules that keep scenarios honest:
- Write the assumption next to the line. "Worst case" alone means nothing. "A wet summer and a competitor cutting prices by 10%" can be argued with, which is the point.
- Do not pretend a scenario is a probability. A scenario is a story about what could happen, not a statement about how likely it is. If you do not know the chance, do not draw it as if you did.
- Keep them few. Three is plenty. Seven scenarios is a fan chart drawn badly.
- Keep the base case equal to the central forecast, so the scenarios and the fan chart tell one consistent story.
Example: three futures for Shannon Outfitters
Sam builds three scenarios for November 2026 to October 2027:
- Base: the central forecast, €1,874k for the year, 5.0% up on the €1,785k of the past twelve months.
- Best: the new lightweight tent range sells as the trial suggested and the shop refit lifts footfall. About 10% above base each month, €2,063k for the year.
- Worst: a wet, short summer (no great stretch of the imagination on the Shannon) plus a competitor's price cut on boots. About 12% below base each month, €1,648k for the year.
The board can now ask useful questions. "In the worst case, does the overdraft cover February?" "What would need to show up by April to tell us the best case is coming?" The fan chart says how wide to plan. The scenarios say what to watch for.
Words for probability: say the number as well
People read probability words very differently. "Likely" means 60% to one person and 90% to another, which is covered in detail in showing uncertainty honestly. The Intergovernmental Panel on Climate Change (IPCC) solved this for its own reports by fixing the words to ranges: "likely" means a 66% to 100% probability, "very likely" 90% to 100%, "about as likely as not" 33% to 66%.9
You do not have to adopt the IPCC scale, but you should borrow the habit: when you use a word, put the number next to it, at least the first time.
Example: the sentence under the fan chart
Sam's first draft said: "Sales will probably reach around €1.87m next year." Aoife has no idea whether "probably" means 55% or 95%.
The rewrite: "The central forecast is €1.87m, 5% up on this year. In any single month, there is a 90% chance sales land inside the light band. Summer months are the least certain: June could reasonably be anywhere from €153k to €209k."
The second version is one sentence longer, and it no longer contains a single word that could mean something different to each reader.
The cone of uncertainty: a misreading to design against
Even a well drawn band can be misread. The clearest evidence comes from weather forecasting. The National Hurricane Center in the United States shows the likely track of a storm as a cone. The cone is built so that about two thirds of the centre's past forecast errors fall inside it, and it widens from about 25 nautical miles at 12 hours ahead to about 200 at 120 hours, in the 2026 Atlantic version.10 It is a picture of uncertainty about where the centre of the storm will go, nothing more.
Members of the public read it differently. In experiments, people made different judgements about the same storm depending on how its uncertainty was drawn,11 and a follow up study found that the visual outline of a summary cone can be misunderstood as showing the size of the storm, as if the hurricane itself grows as it travels.12 Showing a set of possible individual tracks instead, called an ensemble, led to judgements closer to the real probabilities, although people then tended to give too much weight to single tracks.12 A wider review of how uncertainty is shown to the public reaches a sober conclusion: there is limited experimental evidence on how people read these graphics, and what works depends on how comfortable the audience is with numbers.13
The business lesson: a widening fan can be read as "sales will become more volatile" or "the business will get bigger and wilder", rather than "we know less about the far future". The shape is the same. The meaning is opposite.
Example: Aoife's question, and Sam's fix
Aoife looks at the fan chart and asks: "That's grand for November, but why are sales going to swing so much more next autumn?" She has read the band as the size of future swings, exactly like the hurricane cone.
Sam makes three changes:
- the band labels now say what they are: "90% of outcomes expected in this range" instead of "90% interval",
- a one line note on the chart says "The band widens because next autumn is further away, not because sales will swing more",
- in the appendix, a second version of the chart shows 20 simulated possible years as thin grey lines over the central forecast, which, like the hurricane ensembles, shows the idea of many possible futures without a shape that looks like a growing object.
The fan stays the main chart. The fixes make it hard to misread.
Show the track record afterwards
A forecast chart earns trust slowly, and the only way to earn it is to show how past forecasts did. Put last year's forecast, its band, and what actually happened on one chart. Then report two numbers: how far off the central forecast was on average, and how often reality landed inside the band.
The common measure of the first is the mean absolute percentage error (MAPE):14
Where is the number of months compared, is the actual value in month , is the forecast for that month, and the vertical bars mean "ignore the sign", so an error of €5k too high counts the same as €5k too low. Foundations: percentages explains percentage errors from the ground up. MAPE has known weak spots: it breaks down when actual values are zero or near zero, and it punishes forecasts that are too high more than ones that are too low.14 For a business with steady monthly sales, like Shannon Outfitters, it is a fair and easily explained choice.
The second number is coverage: the share of months that landed inside the 90% band. Over twelve months you would expect about inside. Far fewer means the band is too narrow, which, as the M4 results showed, is the usual failure.3
Example: how last year's forecast did
A year ago, Sam forecast November 2025 to October 2026. Here it is against what happened.
Fair play to the model on accuracy. The monthly percentage errors were 6.2, 2.3, 6.6, 0.9, 8.5, 3.4, 7.1, 1.2, 6.4, 5.1, 2.3 and 6.0. Step by step:
Here and the twelve percentage errors add up to . Coverage: 11 of 12 months inside the 90% band, against 10.8 expected. That is about right.
But the chart shows something both numbers hide: every single month came in above the forecast. The model was not just noisy, it was biased, because the business grew faster than the model assumed. A MAPE of 4.7% sounds comfortable. Twelve misses in the same direction is a message: adjust the growth assumption. That is why Sam shows the chart and not only the score.
Putting it together: the forecast page for the board
Everything above fits on one page. Here is the order that works, because it moves from the answer to the doubt to the evidence:
I start every analytics project from the decision it should support rather than from a dashboard, and a forecast page is no exception. If the decision is stock, show months. If it is a bank loan, show the year and the worst case. If it is hiring, show the quarter when the money arrives.
Example: Sam's final page
The decision: how much summer stock to order in January, and whether the overdraft needs extending.
- Fan chart, captioned: "Central forecast €1.87m for the year, 5% up. The band widens because later months are further away, not because sales will swing more."
- Scenarios: best €2.06m (tents and refit succeed), worst €1.65m (wet summer and price cut), each with its assumption.
- The sentence: "Plan summer stock for the central case of €181k in June, but keep supplier terms that let us cut 15% by April. There is roughly a one in twenty chance June comes in below €153k."
- Track record: last year's MAPE 4.7%, 11 of 12 months inside the band, every month above forecast. "This year's model raises the growth assumption to correct the low bias."
Aoife's board does not need to know what exponential smoothing is. They can see the most likely path, how wide to plan, what would change the picture and how much the forecaster has earned their trust.
Conclusion
A forecast is a distribution, so draw it as one. Replace the single line with a fan chart: a central forecast and two bands, 50% and 90%, that widen with the horizon because errors pile up. Know where your band comes from: a time series model gives it for free, a regression gives a band that widens away from the data you have seen, and a machine learning model often gives nothing until you add quantile models, backtest errors or conformal prediction. Assume the band is too narrow rather than too wide. Use a handful of scenarios, each with its assumption written next to it, for the specific risks the fan cannot show. Put numbers next to probability words. Label the band so nobody reads it as the size of the future. And show last year's forecast against what happened, because a track record is the only thing that turns a chart into something a board can trust.