What a Weather Forecast Does for Sales
A weather forecast can be scored every day, and the best are reliable. The sales response is harder: it varies by store and product, good-weather gains are largely repaid within weeks, and the forecast helps a sales model only inside about a week.
Answer to the question
How do retailers forecast sales from the weather?
They add weather variables to a statistical sales forecast and estimate the response store by store and product by product. The weather forecast is the reliable input. The sales response varies by location and product, good-weather gains are largely repaid within weeks, and the forecast adds accuracy only up to about seven days ahead.
Key takeaways
- Allan Murphy and Robert Winkler concluded in 1977 that weather forecasters can give reliable probabilities for rain and temperature, which makes the weather the dependable half of the problem.
- In 673 stores, Florian Badorf and a co-author found in 2020 that the weather effect reached 23.1 per cent across locations and 40.7 per cent across sales themes, and that weather forecasts improved daily sales forecasts up to seven days ahead.
- Brigitte Roth Tran found in 2022 that good-weather gains at one brand's stores were offset by weaker sales in the following weeks, while a very bad day cost a net 40 per cent of a day's sales.
- A 2017 study led by Sebastian Steinker cut forecast errors at a European online fashion retailer by 8.6 to 12.2 per cent on average, and by up to 50.6 per cent on summer weekends, by adding weather data.
"We have to pray for good weather" is not a recommendation anyone can act on. Ethan Rasiel, in The McKinsey Way, sets against it "We must reduce our vulnerability to changes in the weather".
The example is invented, and the book offers no evidence that the practice works. It does set the job for a weather forecast, which is one way a retailer reduces its exposure. An earlier piece, What the Weather Does to Your Spending, asked why weather moves what shoppers buy. This one asks how much exposure a forecast removes. Four studies carry most of what follows. Three were read here at the level of their published abstracts, and the Federal Reserve Bank of San Francisco letter in full.
| Study | Data | Size of the weather effect | Carry-over or horizon |
|---|---|---|---|
| Starr-McCluer, 2000 | American retail sales, monthly | Modest but significant (no figure in the abstract) | Largely gone over a quarter |
| Badorf and co-author, 2020 | 673 stores, daily sales | Up to 23.1% by location, 40.7% by sales theme | Forecast gain up to seven days ahead, shrinking after |
| Roth Tran, 2022 | More than 100 stores of one brand | Very bad day, net loss of about 40% of a day's sales | Good-weather gains offset within four weeks |
| Steinker and colleagues, 2017 | Online orders, a European fashion retailer | Good weather cut orders, bad weather raised them by less | Forecast errors down 8.6 to 12.2% on average, 50.6% at most |
How reliable is a weather forecast?
In 1977 Allan Murphy and Robert Winkler summarised American programmes in which forecasters gave probabilities for rain and for the day's highest and lowest temperatures. Their conclusion was that weather forecasters can make such forecasts reliably. Reliable has a precise meaning here: of the days given a thirty per cent chance of rain, about three in ten turn out wet. The weather resolves every day, so the forecast can be scored every day.
A forecast can be reliable without being sharp. A forecaster who gave the average chance of rain on every day would be reliable and of no use, which is why the retail studies below measure how much forecast error falls, not reliability. A retailer starting from a forecast like that has one trustworthy input. The next step has to be measured separately: how many units a given temperature moves.
The weather resolves every day, so the forecast can be scored every day.
Modern weather forecasts also lean on ensembles, in which a model is run many times from slightly different starting points and the spread of the results shows how uncertain the day is. Peter Bauer and his co-authors name ensemble forecasting as one of four pillars of numerical weather prediction in their 2015 review in Nature, alongside observations, the description of physical processes and the computing power to run them. Noise makes a related point about forecasts in general: it reports that unweighted averages of forecasters beat most individuals, and sometimes all of them, in sales, weather and economic forecasting (p. 261).
How far do sales respond?
A 2020 paper by Florian Badorf and a co-author measured the response on daily sales from 673 stores, and found it large and uneven. The effect reached 23.1 per cent across locations and 40.7 per cent across what the paper calls sales themes. Direction varied as well as size, so one warm day could lift one range of goods and cut another. Models that treated the response as a straight line misjudged extreme weather. The method is the reason the paper can say so: a random coefficient model gives each store its own response instead of imposing one average, so the same weather can be seen doing different things in different places.
The best-known illustration is a single sentence. Linda Dillman, Wal-Mart's chief information officer, told The New York Times in 2004 that strawberry Pop-Tarts sell at about seven times their normal rate ahead of a hurricane, and that beer was the top seller before one. No data were shown, but the statement locates the signal: particular products ahead of a particular event, not the total at the till.
How much of the response comes back?
Some of the gain does. A warm weekend can lift a week's sales without adding to a month's, if the sales were only brought forward. Martha Starr-McCluer tested this for American retail in a 2000 Federal Reserve paper. Her model treats weather as a change in how productive people's time is for shopping and recreation, which can move spending between goods and between periods. Unusual weather had a modest but significant effect on monthly sales, later months often offset the first effect, and over a quarter the influence had largely gone.
Daily figures show the same shape, with an asymmetry, in a 2022 letter by Brigitte Roth Tran. Her data were daily sales at more than 100 stores of one national apparel and sporting goods brand between 2010 and 2013. Good-weather gains were offset by lower sales in the following weeks and were no longer statistically significant by four weeks. Bad-weather losses persisted. A moderately bad day cost a net quarter of a day's sales, and a very bad one about 40 per cent.
Two other channels might have absorbed the loss and did so only a little. Online sales did not offset the bad-weather losses in stores, and shoppers moving between indoor and outdoor shopping centres offset only a small share of them. Experience also changed the response.
Stores in places with more history of a kind of weather lost less to it: for precipitation over two inches, the average sales decline was about 80 per cent smaller where experience was above the median. A model fitted in one climate will misjudge shops in another, and Roth Tran warns that weather arriving in places that rarely see it could make today's average responses understate the disruption.
On those figures, a forecast of a warm weekend should raise this week's order and trim the next few, while a forecast storm should cut the order and nothing should be pencilled back in. From one brand's stores, that is a pattern to test. Starr-McCluer's national monthly series and Roth Tran's daily store sales are different objects, and their results should not be set against each other as if they measured the same thing.
How far ahead does the forecast help?
Badorf and his co-author added weather forecasts to the daily sales forecast and found accuracy improved up to seven days ahead, with the benefit shrinking beyond that. Online, the weather behaves differently.
A 2017 study led by Sebastian Steinker used about two and a half years of orders from a European online fashion retailer. Good weather cut orders strongly and bad weather raised them by much less, with larger effects in summer and at weekends. Adding weather data to the forecasting models, compared with a baseline time-series model that ignored it, cut forecast errors by 8.6 to 12.2 per cent on average, and by up to 50.6 per cent on summer weekends.
The same paper tested workforce planning at an order fulfilment centre. A scenario in which the weather was known perfectly cut excess costs by 11.6 per cent against a baseline that ignored it. Nobody knows next week's weather perfectly, so 11.6 per cent is the most that information could be worth.
The opposite sign appears in Roth Tran's data, where the online sales of the brand she studied rose on good-weather days. The retailers, countries and years differ, and neither study can say which difference matters. How large the online channel is, and how its own forecasts have fared, is the subject of Why Online Shopping Didn't Take Over.
Whether the buyers who place the orders take the model's number is a separate question, covered in How Much Advice People Actually Take. The abstracts do not say whether any of these studies set the model against a buyer's own judgment.
Applied to a retailer, the studies support estimating the weather response from its own sales, store by store and range by range, because the response differs by store, by range and by climate. They support using the forecast inside the week, where it has been shown to add accuracy, and treating a good-weather lift as mostly borrowed from the following weeks and a bad-weather loss as mostly gone, while a figure that another firm measured on different shops, in a different climate and range, does not transfer to the retailer's own stores.
Common questions
How do retailers forecast sales from the weather?
They add weather variables to a statistical sales forecast and estimate the response store by store and product by product. In a 2017 study led by Sebastian Steinker, adding weather data cut forecast errors at a European online fashion retailer by 8.6 to 12.2 per cent on average.
How far ahead does a weather forecast help a sales forecast?
About seven days. Florian Badorf and a co-author found in 2020 that weather forecasts improved daily sales forecasts up to seven days ahead, with the benefit shrinking at longer horizons.
Is the weather forecast the reliable part of a sales forecast?
Yes, as far as the evidence goes. Allan Murphy and Robert Winkler concluded in 1977 that weather forecasters can give reliable probabilities for rain and temperature. The harder step is the sales response, which differs by store and by product.
Does good weather always raise retail sales?
No. In Brigitte Roth Tran's 2022 data, gains on good-weather days were offset by weaker sales in the following weeks, while losses on very bad days cost about 40 per cent of a day's sales and persisted.
Sources
Murphy, A. H., & Winkler, R. L. (1977). 'Reliability of Subjective Probability Forecasts of Precipitation and Temperature.' Journal of the Royal Statistical Society, Series C (Applied Statistics), 26(1), 41-47. doi:10.2307/2346866
Badorf, F., & Hoberg, K. (2020). 'The impact of daily weather on retail sales: An empirical study in brick-and-mortar stores.' Journal of Retailing and Consumer Services, 52, 101921. doi:10.1016/j.jretconser.2019.101921
Steinker, S., Hoberg, K., & Thonemann, U. W. (2017). 'The Value of Weather Information for E-Commerce Operations.' Production and Operations Management, 26(10), 1854-1875. doi:10.1111/poms.12721
Starr-McCluer, M. (2000). 'The Effects of Weather on Retail Sales.' Finance and Economics Discussion Series 2000-8, Board of Governors of the Federal Reserve System.
Roth Tran, B. (2022). 'The Impact of Weather on Retail Sales.' FRBSF Economic Letter 2022-23, Federal Reserve Bank of San Francisco, 22 August 2022.
Hays, C. L. (2004). 'What Wal-Mart Knows About Customers' Habits.' The New York Times, 14 November 2004. Dillman's statements read in a reposted copy of the article.
Rasiel, E. M. (1999). The McKinsey Way. McGraw-Hill, pp. 10-11.
Bauer, P., Thorpe, A., & Brunet, G. (2015). 'The quiet revolution of numerical weather prediction.' Nature, published 3 September 2015. doi:10.1038/nature14956
Kahneman, D., Sibony, O., & Sunstein, C. R. (2021). Noise: A Flaw in Human Judgment. Little, Brown Spark, p. 261.