Today I spent more time comparing forecasting models for sales data, and honestly, this is one of the most interesting parts of building prediction software. At the beginning, sales prediction sounds simple. You imagine giving the system some numbers and asking: "What will I sell next month?"
But when you actually start working with real sales data, you quickly understand that sales are not clean, polite or predictable. Some products sell every day. Some products disappear for weeks and then suddenly sell again. Some follow seasons. Some react to discounts. Some depend on stock, weather, holidays or simple human behavior.
And then you realize something important: there is no one magic model.
One of the models I tested is Seasonal Naive. It is almost funny how simple it is. It basically says, "Look at the same period in the past and use that as the prediction." Last Monday. Last December. Last summer. Nothing fancy. But this simple model is very useful, because it keeps you honest. If a more advanced model cannot beat Seasonal Naive, then maybe the advanced model is just decoration.
Then I looked at Croston and SBA, and this is where things became more interesting. These models are made for products that do not sell all the time. The kind of product that has many zero-sales days, but when demand appears, it matters. Instead of panicking because there are many zeros, Croston-style models try to understand two separate things: how often the product sells, and how much it usually sells when it finally moves. SBA is a more corrected version of that idea. For slow-moving products, spare parts, luxury items or anything with irregular demand, this makes a lot of sense.
Then there is LightGBM with Tweedie loss, which feels like bringing a much more serious machine to the table. This model can learn from many signals together: previous sales, prices, promotions, seasonality, product category, store, holidays and other business data. The Tweedie part is very useful because sales often have a strange shape: many zeros, but sometimes big numbers. That is real life. Not every product sells smoothly.
What I find fascinating is that each model has its own personality. Seasonal Naive is simple but honest. Croston and SBA are patient with irregular demand. LightGBM Tweedie is more powerful and can understand many business signals at the same time. And this is exactly what makes sales prediction difficult, but also exciting.
A coffee product, a spare part, a hotel service, a medicine, a clothing item and a luxury product do not behave the same way. So the software should not force one model on everything.
My approach is to let the system compare models, test them on past data, and choose what works best for each product or category.
Because the goal is not to use the most impressive algorithm. The goal is to help a business make better decisions. It is business reality, translated into signals.
Originally published on LinkedIn.