Delivery time prediction is the part of a food delivery app that calculates the number of minutes it shows you before you order, such as "Delivery in 32 minutes". This number is called the estimated time of arrival, or ETA. The app must then deliver within that time, so it also tracks each order against the promise.
The prediction adds up three times. The first is the kitchen's cooking time: how long this restaurant usually takes to prepare this kind of order. The second is the time the rider needs to reach the restaurant. If the rider arrives before the food is ready, this time adds nothing, because it overlaps with the cooking. The third is the ride to your door, including the traffic and the minutes spent finding your building.
Each of these times is predicted by a model, a program trained on millions of past orders. The model looks at the same restaurant at the same hour on the same day of the week. It also looks at the weather, the number of free riders and the live traffic on digital maps.
Here is a typical case. On a rainy Friday at 8 p.m., a student in Pune looks at a pizza restaurant 4 km from her hostel. The model expects the kitchen to take 16 minutes, because this restaurant's pizzas take longer on Friday evenings. The nearest free rider is 5 minutes from the restaurant, so he will be there before the pizza is ready. The 4 km ride is expected to take 14 minutes in the rain. Finding the hostel gate and the right block adds 2 minutes. So the app shows 32 minutes.
The tracking starts after she orders. At the tenth minute, the kitchen marks the pizza as 5 minutes late. The app changes the time to 37 minutes and tells her. Route optimisation software can send the rider a little later, so that he does not stand waiting at the counter.
Why does the number matter so much? If the app promises too few minutes, orders arrive late, customers complain and the app may have to give refunds. If it promises too many, customers order from another app. So the app compares every promise with the real delivery time, and trains the model again.
The model also needs to know how many riders will be free in each area. That comes from demand forecasting. The same kind of prediction is used in quick commerce, where the promise is about ten minutes, and in a ride-hailing app, which shows when your cab will arrive.