External events

How to incorporate them into a sales forecast

What an exogenous event really changes in a forecast

A sales forecast is based primarily on internal signals: sales history, price, promotions, and recurring seasonality within a category. These are endogenous signals—they come from the sales system itself—and a good model almost always captures them accurately if it has a clean and sufficient historical record, which is the subject of another article in this series.

An exogenous event, on the other hand, originates from outside the sales system. It isn't reflected in the historical data—unless a similar event has already occurred—and it can throw off an otherwise solid forecast.

Four categories account for the vast majority of retail scenarios: weather, the retail calendar (holidays, school breaks), current events and trends, and a competitor's actions.

First instinct: When an external event is detected—such as favorable weather boosting projected demand or an upcoming holiday—the question isn’t just “how much to order,” but also “should we adjust the price or promotion accordingly?” An event that is accurately anticipated by the forecast but does not result in any pricing decision makes no difference to the end customer: price remains the primary lever available to influence their response.

The Booper guide on the mechanics of forecasting— AI-Driven Sales Forecasting: Methodology and KPIs —mentions these factors in a single sentence, listing them among the biases and pitfalls to watch out for. This is intentional: that article focuses on the formula, the data, and the accuracy KPIs, not on the details of each external factor.

This article picks up where the previous one left off on this specific point, raising a key question that the earlier article did not address: Not all exogenous events are equal, so which ones truly deserve to be modeled?

The weather, the primary external factor—and the most well-documented

Weather is the most thoroughly studied external factor—and the one most easily underestimated until it has been quantified.

This figure combines effects that vary greatly by category—heavy rain does not affect the garden section in the same way as it does the dry goods section. An academic study published on ScienceDirect, which focuses specifically on brick-and-mortar stores, goes a step further by quantifying the increase in accuracy that comes from incorporating weather data into a forecasting model.

In practical terms: temperature, precipitation, and sunshine have a major impact on the garden, barbecuing, ice cream, cold drinks, seasonal clothing, and outdoor DIY projects. They have almost no impact on dry goods, home maintenance, or daily personal care and beauty routines.

Incorporating weather data everywhere, without distinction, amounts to making a model more complex without any benefit for a large portion of the dataset—and that is precisely what the prioritization method described below helps to avoid.

Holidays, school breaks: The calendar takes its toll, but not everywhere in the same way

The second major category of exogenous events is calendar-related:

• Fixed holidays (May 1, July 14, Christmas).
• Movable holidays tied to Easter (Easter Monday, Ascension Day, Pentecost).
• School holidays by region.

What these events have in common is that they are known in advance, which, in theory, makes them the easiest to account for—and yet many forecasting models struggle with them, because a movable holiday never falls on the same date from one year to the next and disrupts seasonality based on the calendar year.

The same INSEE Focus report shows that other sectors (bakeries and pastry shops, fishmongers, and home appliances) experience a year-end peak that is significantly more pronounced than the retail average, while others have almost none. It’s the same principle as the weather, applied to the calendar: the external factor is real and measurable, but its impact depends entirely on what you sell.

School breaks: a more localized impact

School breaks add another, more local dimension: a store in a tourist area or resort sees its foot traffic shift with Zones A/B/C, regardless of any national holidays. A downtown store in a major city, on the other hand, may see its customer traffic drop during summer break, when the local clientele leaves.

The same calendar event has two opposite effects depending on location—yet another reason to prioritize by category and by area, rather than uniformly across an entire network.

Current events, viral trends, and competitor actions: signals that no historical data contains

The last two categories of exogenous events have one thing in common: by definition, they have never occurred in the same way before.

• Current events and viral trends. A topic that’s trending on social media, a recipe or product that goes viral, or a product recall by a competitor that shifts demand elsewhere: none of these signals appear in any sales history before they first occur. A traditional statistical model literally cannot anticipate them—only active monitoring, whether done by humans or automated tools, can detect them in time to adjust an order or a price.
• Competing actions. Opening a store nearby captures a portion of existing foot traffic; closing one frees up another portion. An aggressive promotional campaign by a competitor in a given category temporarily shifts demand, before things return to normal once the campaign ends. These are local signals, often limited in time, that require continuous monitoring rather than a fixed parameter in the model.

These two families are not modeled in the same way as weather or calendar data: they are monitored using an alert system capable of flagging deviations before they cause lasting distortion in the demand forecast.

Of these two categories, actions taken by competitors have the most direct impact on price: a competitor’s store opening, closing, or aggressive promotional campaign is an exogenous event that almost always affects the optimal price of an item—not just the volume of sales.

A category manager who accurately forecasts the volume impact of a competitor’s promotion but fails to adjust their own price in response misses out on an opportunity—or exposes themselves to the opposite risk. This is also where the topic intersects directly with promotional activities: a poorly calibrated price response to a competitor’s promotion can cost more than the promotion it is intended to counter.

The method for prioritizing: Not all exogenous events have the same impact

Faced with four families of factors and a catalog of several thousand items, it’s tempting to add all available variables to the model “just in case.” This is the most common mistake: a model overloaded with weakly correlated factors becomes harder to interpret, without gaining accuracy where it counts.

The right approach is to prioritize before integrating.

Prioritization Grid by Product Category

This framework provides a starting point for prioritizing the integration of the four categories of exogenous events by product category. It does not replace a correlation test on your own data (Step 1 above), but it does prevent you from having to start from scratch.

To put it this way: the dry goods category has virtually no “strong” exogenous factors—it’s a category where it’s best to focus modeling efforts on historical data and promotions. The garden and toys categories, on the other hand, each have a different dominant factor (weather for one, the calendar for the other), which justifies a dedicated approach.

This matrix can also be interpreted in terms of price, not just inventory: a “Strong” factor—such as the weather affecting garden products or the holiday season affecting toys—justifies price or promotional adjustments that are just as responsive as inventory adjustments. Conversely, a category where everything remains “Low” or “Medium”—such as dry goods—generally does not require event-driven pricing: standard pricing and promotional mechanisms are sufficient.

Incorporate these signals without unnecessarily complicating the model

Once the priority factors have been identified for each category, technical integration follows a simple hierarchy:

• First, there are calendar-related factors (holidays, school breaks), which are known months in advance and can be automated without any particular effort.
• Next, the weather forecast, using a 7-day or 14-day forecast stream, depending on the decision horizon.
• Finally, monitoring signals—news, trends, competitor actions—which are more akin to an early-warning system than to a fixed variable in the statistical model.

The pitfall to avoid is the opposite of what one might imagine: it’s not including too few exogenous factors, but including too many—without establishing a hierarchy—to the point where the model’s forecasts become impossible for anyone to explain. A model that clearly explains why it forecasts a 12% increase in the gardening department next week (favorable weather forecast, no holidays, no competing promotions detected) is more useful for decision-making than a more complex but opaque model.

This is also a prerequisite for maintaining a useful sales history over time—a topic covered in detail inthe article in this issue dedicated to sales history.

Key Takeaways

External events are not an optional detail in sales forecasting, but neither should they all be treated with the same level of attention. Weather can account for up to 56% more variance in one sensitive category, yet virtually none in another.

The holiday shopping season accounts for 27.9% of annual sales for a toy store in December, compared with an average of 10% for the retail sector.

Current events and competing actions, on the other hand, cannot be modeled—they must be monitored. The right approach does not simply add up all available factors: it prioritizes them, category by category, before integrating them.

The common thread running through this article: gardening in the heat, pre-Christmas toys, and responding to a competitor’s promotion—in all three cases, the question of price is just as important as that of how much to order. These factors are all the more valuable when they are integrated into a single, modular pricing platform, rather than into a forecasting tool that is isolated from other pricing and inventory decisions.

Want to know which external factors really matter for your catalog?

Spend 30 minutes with our team to identify, category by category, what deserves to be included in your forecast—and what isn't worth the effort.

Let's plan an exchange →

Frequently Asked Questions

Also in this series

• Sales Forecasting: Methods, AI, and Best Practices (Feature Article)
• What sales history is needed to make accurate forecasts?

Sources

• American Meteorological Society, as reported in several academic studies (including those by the Federal Reserve Bank of San Francisco and ScienceDirect)—more than 3% of retail sales are directly affected by the weather, amounting to more than 1,000 billion dollars worldwide.
• ScienceDirect, " The Impact of Daily Weather on Retail Sales: An Empirical Study in Brick-and-Mortar Stores " — up to an additional 47% of variance explained for a product and 56% for a category by incorporating weather into the forecasting model.
• INSEE Focus No. 170, Holiday Season, Sales, “Black Friday”: A Significant Impact on Retail Sales, November 26, 2019 — insee.fr
• Booper, AI-Powered Sales Forecasting: Methodology and KPIs — booper.fr/blog
• Booper, internal product data (`context/socle_booper.md` §3) — GENIUS Predict, GENIUS Monitoring, GENIUS Price, Pricing Analytics.