Time Series Forecasting

We use time series forecasting to develope a model that can predict the future values of a time series based on previous data. The length of time into the future for which the forecast is prepared is called a forecast horizon.

If the values of either a time series or the function x(t) (where t = time) are numbers, then you're facing a regression task for the time series. If they are categories, it will be a classification task.

Using the initial data, we will create a training set and a test set. You can't mix the sets in time series forecasting, and the training set data must precede the test set data. Otherwise, model testing will be flawed: the model shouldn't be trained on future data after all. The train_test_split() function from the sklearn.model_selection module mixes data by default, so let's set the shuffle argument to False so the data can be separated correctly into training and test sets:

COPYpython
1import pandas as pd
2from sklearn.model_selection import train_test_split
3
4train, test = train_test_split(data, shuffle=False, test_size=0.2)

Forecast Accuracy

We'll be using the MAE metric to evaluate the models in our tasks because this metric is easily interpretable.

There are two ways to forecast time series without training:

1. All values of the test sample are predicted with the same number (a constant). For the MAE metric, this number is the median.
2. The new value x(t) is predicted by the previous value in the series, defined as x(t-1). This method is metric-independent.

Creating Features

1. Calendar features

Usually trends and seasonality are linked to a specific date. The datetime64 type in pandas already contains the necessary information, and all that remains is to present it as separate columns. Let's consider this example:

COPYpython
1# this feature contains years as numeric values
2data['year'] = data.index.year
3
4# this feature contains weekdays as numeric values
5data['dayofweek'] = data.index.dayofweek

2. Lag features

The previous values in the time series will tell you whether the function x(t) will grow or decrease. Let's get the lag values using the shift() function:

COPYpython
1data['lag_1'] = data['target'].shift(1)
2data['lag_2'] = data['target'].shift(2)
3data['lag_3'] = data['target'].shift(3)

Not all lag values are available for the first dates, so these lines contain NaN.

3. Rolling mean

The rolling mean feature sets the general trend of the time series. Here's a reminder of how to calculate it:

COPYpython
1data['rolling_mean'] = data['target'].rolling(5).mean()

The rolling mean at t takes into account the current value of the x(t) series. This is incorrect: target "slipped" into the features. The rolling mean calculation should not include the current value of the series.