In this post, let us explore the basic concepts about time series. We will also learn about resampling techniques, how to check for stationarity and ways to convert non stationary series into stationary series.

**What is time series**

In time series, data are recorded over time. Time interval may be daily, monthly, yearly etc.

**How to import time series data**

I have written about importing time series data into python in this post.

**Handling missing values**

This post provides information on handling missing values. Depending upon the case, generally it is advisable to fill the missing values with previous value or next value or using interpolation techniques or moving averages or using time series models.

## Resampling:

## A) Downsampling

In simple terms, it is like aggregating. For example: converting daily data to monthly data, or quarterly data to yearly data etc.

In the following example, I have converted daily data to weekly data.

Original data

In resample option, use rule=’W’ for weekly frequency. Other commonly used rules are B (Business day), M (Monthly), A or Y (year end frequency). For more rules, you can refer this document.

In the below picture, not how changing the label=’left’ changes the week in resampled data3. We can aggregate the data using measures such as ‘sum’, ‘mean’ etc.

## B) Upsampling

Here we increase the frequency in time series data. It is the exact opposite of downsampling. For example resampling the data from weekly to daily data, or converting yearly data to quarterly data. But increasing frequency leads to empty values. Empty values can be handled in the similar way as we handle missing values in time series data.

In the following upsampling example, let us convert non continuous data to ‘business day’ frequency and fill the empty values with bfill method.

Original data

Upsampled data with ‘bfill’ method to handle empty values

**Visualizing time series data**

To plot:

plt.plot(data)

## Stationarity

There are two types of stationarity:

- Strict stationarity : all time series properties are time invariant.
- Weak (Wide-sense stationarity): mean, variance and autocorrelation are time invariant.

In simple terms, a time series whose mean and variance do not change over time, is said to be stationary (more about stationarity)

## Why we need the series to be stationary

If we want to predict future, then if we fit a good model on stationary data, it will accurately predict the future as its statistical properties do no change over time.

## How to test the stationarity

**Plots-**Visualization will give a broad idea about trend, variation in the time series**Augmented Dickey–Fuller test**

Alternate Hypothesis: Series is stationary.

Augmented Dickey–Fuller test

For this test, I have used the code from this blogpost and slightly modified it.

**Kwiatkowski-Phillips-Schmidt-Shin Test**

## a) Null hypothesis that time series is level stationary

## b) Null hypothesis that time series is trend stationary

Additionally use regression =’ct’ option to test trend stationarity.

KPSS test with Null hypothesis-trend stationarity

Phillips–Perron test can also be used to test stationarity.

## How to convert non stationary series to stationary series

**Log transformation**

**Differencing**

**Seasonal differencing**

Now perform seasonal differencing (since this is the monthly data, 12 will be the order of seasonal differencing).

We can use any combinations of these transformations to transform the time series depending upon the situation.

## White Noise

White noise is a series whose values are uncorrelated with zero mean and finite variance. We want the error terms of our model to be a white noise process. It is a stationary series.

## Random walk

If an observation is dependent only on its previous value(as shown below), then that series is said to follow random walk.

Y_{t }=Y_{t-1}+Et

which is,

Y_{t }-Y_{t-1 }= Et (which means differenced series is white noise)

Another type of random walk model: random walk with drift (α).

Y_{t }= α+Y_{t-1}+Et

Y_{t }-Y_{t-1 }= α+Et (which means differenced series is equal to white noise + α)

One thing to note that while random walk series is non-stationary, not all non-stationary series follow random walk (more on this).

## Summary

In this post, we have explored

**what is time series data****resampling (downsampling and upsampling)****stationarity-concept and tests****differencing****white noise****random walk model**