Time series Analysis

Ever wondered how it would feel like to predict the future?

1. We first started making predictions using the naive approach, which considers the last value of the data set as the future value and the results were awful, as you can see below.

2. Next, instead of taking just the last value, we averaged the last 10, 20, and 50 values. The forecast was just a bit better!

3. To improve things further, we decided to assign weights to the records, with higher weights to more recent records and lower weights to older records. We saw a bit more progress!

4. Now, we somewhat understood the concept of forecasting and wanted to involve more statistics. We implemented Holt’s linear model which considers the trend of time series data to make predictions.

5. We then realized that there was seasonality in the data, rather then a trend, so we implemented another stats model called Holt Winter’s model. And BOOM! It showed some silver lining!

6. Now that we knew that we needed to leverage the fact of seasonality of our data in these statistical models, we thought of trying the ultimate statistical time series forecasting model called SARIMA. We had to play a lot with the parameters, but the results were great, as you could see below.

7. Finally, we tried to use this model to predict the traffic for the month of December 2018 on daily basis and below is the result!

Note: The reason why the model above starts at December 5th and ends on December 30th is because we ran the model on the 5th.

If you want to learn more about how this was done, then check out the code on GitHub.