Maytime is an automation agency for growing businesses without an in-house technical team.
We look at the work your team does by hand and build automation around the tools you already use.
, it can track whether interest in a category is beginning to move before the same change appears clearly in sales. These events are not future purchases. They are early evidence. The important information is not simply that searches increased. It is whether they are increasing **earlier or faster than would normally be expected**. ## Looking outside the retailer The retailer's own data should be the starting point, but external signals can help answer a different question: **Is this happening only in our store, or is interest in the category changing more widely?** Search data can help here. Google Trends, for example, can show whether interest in a category is accelerating relative to its normal seasonal pattern. One study covering [5,577 products sold between 2016 and 2019](https://arxiv.org/abs/2109.09824) found forecasting performance improved when Google Trends information was added. That does not mean Trends automatically improves this retailer's forecast. It makes it something worth testing. Other useful signals could eventually include wider product-search behaviour, social interest or category-level trend data. Each new source should have to prove that it improves the warning. More data is not automatically a better forecast. ## Weather works differently over this time frame Weather matters for products such as paddling pools, but there is an important limitation. A reliable forecast for a specific hot weekend several months away does not exist. So the system should not pretend otherwise. Over a longer horizon, weather becomes supporting context rather than a precise prediction. The system might use broader seasonal outlooks alongside historical weather relationships to understand whether conditions are becoming more or less favourable for a category. Then, as the season gets closer, shorter-range forecasts become more useful. The prediction can therefore change as new information arrives. Eight weeks out, historical seasonality and early customer behaviour might matter most. Three weeks out, search momentum and current sales become stronger signals. Seven days out, actual weather forecasts can carry much more weight. The forecast should become more specific as uncertainty falls. ## The forecast should update continuously This is not a prediction made once in January and forgotten about. The system keeps comparing what is happening with what it expected to happen. Suppose paddling-pool interest starts rising unusually early. The system increases the expected demand range. Two weeks later, searches continue accelerating and sales begin following. Confidence increases again. Or the opposite happens. Interest rises briefly, then disappears. The forecast falls back towards the normal seasonal range. This matters because the retailer does not have one stock decision. It has a series of them. The useful question changes from: **What will summer demand be?** to: **Given everything we know today, has our view of summer demand changed enough to alter the next decision?** ## Prediction should follow the purchasing calendar This is where the system becomes useful operationally. Different decisions need different amounts of warning. A retailer might have: * Three months to place an initial seasonal order * Six weeks to increase an incoming shipment * Three weeks to adjust stock between warehouses * One week to change advertising * A few days to change pricing or promotions The prediction does not need the same accuracy at every stage. Three months out, the retailer may only need to know that one category looks unusually strong. Closer to the season, it can make increasingly precise decisions. The forecast becomes useful because it is connected to **what the business can still change**. ## Test it before trusting it The first system should run without changing any orders. It can replay previous seasons first. For each point in the year, the system is only allowed to see information that would genuinely have been available at the time. Then you can ask: **What would it have predicted six weeks before the spike?** **What would it have predicted four weeks before?** **What would it have predicted two weeks before?** This avoids accidentally building a system that only looks clever because it already knows what happened. Once that historical testing is useful, the system can run alongside the buyer for a live season. Its predictions are recorded but purchasing remains unchanged. The retailer can then see whether the warnings genuinely arrived before its existing process would have spotted the change. ## Measure whether knowing earlier changes anything Forecast accuracy matters. But it is not the only measure. A retailer could have a statistically impressive forecast that arrives too late to affect anything. The more useful measures include: * How many weeks of useful warning the system provided * Whether it spotted seasonal movement before sales alone * How often it produced false warnings * Whether it missed important seasonal changes * Whether the warning arrived before a purchasing deadline * How the forecast improved as the season approached Eventually, a live test can measure the commercial result. Did the retailer hold less excess stock? Did it run out less often? Did fewer emergency orders have to be placed? Was less stock discounted after the season? Did purchasing become more confident? That is when the forecast starts earning its place. ## Where this can go wrong Seasonal prediction is uncertain. Something appearing earlier than last year does not mean it will keep growing. Marketing can distort customer behaviour. Online trends can disappear quickly. A category can generate huge interest without generating purchases. Historical patterns can break. Weather can move in the opposite direction. Competitors, prices and wider consumer behaviour can change. There is also a temptation to solve uncertainty by adding more data. That can make the system worse. Twenty weak indicators do not necessarily beat four useful ones. The aim should be to discover the smallest set of signals that consistently tells the retailer something it did not already know. ## This is really about recognising change early The interesting part of this system is not predicting an exact number months into the future. It is recognising when the shape of a season is beginning to change. Is summer interest appearing earlier? Is one category accelerating faster than normal? Is a trend that normally peaks in July starting to build in May? Is this year's demand trajectory beginning to separate from the historical pattern? Those changes can appear weeks before they become obvious in sales. That creates time. Time to change an order. Time to speak to a supplier. Time to move inventory. Time to change what gets promoted. At Maytime, we would start with one category and one upcoming season. Build its normal seasonal pattern. Identify the earliest useful signals. Then test whether those signals could reliably move a real purchasing decision forward. Not to know exactly what the future will look like. But to recognise when it is beginning to look different. ## Start with one season Take one seasonal category in your business. Write down: 1. When its season normally begins. 2. When you have to start committing to stock. 3. Which signals appear before sales begin moving. 4. How those signals behaved in previous years. 5. What you could change with four, eight or twelve weeks of warning. That gap between the first useful signal and the last useful decision is where forecasting becomes valuable.](https://cdn.sanity.io/images/zcx2jd66/production/892d693909a78a8567013ea380f6a576e0a7aede-1254x1254.png?w=960&auto=format)
Our mission is to make the advantage of automation available to every growing business.
Ready to scale your company?
Show us what is slowing your team down. We will find the practical automations that make growth easier.