Uses and Limitations of Time Series

Understanding data

Another benefit of time series analysis is that it can help an analyst to better understand a data set. This is because of the models used in time series analysis help to interpret the true meaning of the data, as touched on previously.

Opportunity to Clean data

The first benefit of time series analysis is that it can help to clean data. This makes it possible to find the true “signal” in a data set, by filtering out the noise. This can mean removing outliers, or applying various averages so as to gain an overall perspective of the meaning of the data.

Of course, cleaning data is a prominent part of almost any kind of data analysis. The true benefit of time series analysis is that it is accomplished with little extra effort.

Forecasting data

Last but not least, a major benefit of time series analysis is that it can be the basis to forecast data. This is because time series analysis by its very nature uncovers patterns in data, which can then be used to predict future data points.

For example, autocorrelation patterns and seasonality measures can be used to predict when a certain data point can be expected. Further, stationarity measures can be used to estimate what the value of that data point will be.

Really, it’s the forecasting aspect of time series analysis that makes it so popular in business applications. Analyzing and understanding past data is all good and well, but it’s being able to predict the future that helps to make optimal business decisions.

Time Series Analysis Helps You Identify Patterns

Memories are fragile and prone to error. You may think that your sales peak before Christmas and hit their bottom in February… but do they really?

The simplest and, in most cases, the most effective form of time series analysis is to simply plot the data on a line chart. With this step, there will no longer be any doubts as to whether or not sales truly peak before Christmas and dip in February.

Limitations

Time series methods draw on vastly different areas in statistics, and lately, machine learning. You have to know a lot about all of these things, in general, to make sense of what you’re doing. There is no real unification of the theory, either.

Often there are ways around getting a model that is time-series based where the predictions are almost as good and is faster to implement. Note that this may or may not blow up in your face later on. In some cases, however, temporal effects are so weak that it makes more sense to just use the non-temporal ones… which can be difficult to explain (the need to check) to a manager if we’ve had to spend 2.5 weeks setting up the tests for temporal effects. Personal experience here.

This is hard stuff, and if you’re not motivated by challenge, you can get overwhelmed. Also, there is, in some other areas of data science, the notion that all we use are ARIMA models and EWMA; while we do often use these tools, we also use RNN and LTSM networks and a whole lot of interesting things.

Most machine learning algorithms don’t deal with time well.