What is the best machine learning model to use for traffic forecasting?

Question

Mathematics

Hello.
I have fixed route traffic data for some time. That is, there is data on the beginning and end of movement along the route during the day for several months. At the same time, the duration of the trip varies during the day depending on traffic congestion, and also depends on the day of the week - roads are freer on weekends and the load is different than on weekdays. In addition, there may be some seasonality - I do not know this yet.
The task is to predict the time of movement along the route, depending on the time of day and day of the week.
What algorithm or machine learning model is used for such cases? And what factors can the model have besides the actual time of day (and, possibly, the day of the week)?
I saw an article that said that machine learning has nothing to do with it at all, but then what?
I would be grateful for any lead.

Answer 1

Hello.
Read ( How a short-term forecast works on Yandex.Traffic ), perhaps new ideas will appear to optimize the idea.
For the development of predictive models in the presence of time series, the most popular models are autoregressive, integrated moving average (ARIMA) (for univariate time series with a trend and without seasonal components) and SARIMA (with trend and / or seasonal components).
Links:
1. statsmodels.tsa.statespace.sarimax.SARIMAX
2. statsmodels.tsa.statespace.sarimax.SARIMAXResults

Answer 2

I see that you do not have enough data for accurate forecasting.
It's like predicting whether a person will be late for work based on statistics about how on time they were all the time before. The maximum that can be done is to identify some system properties that are more or less constant and rarely change. For example, the chance of being late. This is a specific figure that can be calculated and further predicted based on it. But the accuracy of the prediction will be equal to the chance of being late. That is, this very chance will be an error.
Applications that are now engaged in predicting the schedule, when which transport will arrive, have much more information at their disposal:
Of course, the accuracy is still not 100%. And sometimes there are overlaps. For example, a phantom bus may pass by a stop, or, conversely, a bus may arrive that does not exist according to the application. But still, the accuracy of forecasting is much higher than if based only on a retrospective analysis.

Answer 3

First of all, you need to come up with some kind of mathematical model that would describe the duration of the flight.
The classic solution is to take the flights as a Poisson flow. Then you can predict the duration of the flight with any probability. Such a model gives a rough upward estimate, so there should not be any particular problems.
It’s a long story about machine learning, but the idea is simple:
1. classify
2. train
3. predict