Time Series Forecasting in Python: Chapter 02
Disclaimer
The contents here are the notes taken by me while I was reading the book Time Series Forecasting in Python by Marco Peixeiro. The book can be found online at this link. The book is also available at Oreilly online library if your organization has a subscription to Oreilly.
The write up here is merely for notes so that I can revisit later on demand. The rights to the contents of the books remain with the author.
My Notes from Chapter-02: Naiive Forecasting
- Future data can be predicted to the best ability as the past data is available. Thus some statistical function of past data can be used for prediction.
- Option 1: Historical Mean: Taking a historical mean of the available data. This gives a fair idea on expected behavior of the data that might occur in the future. However, this is not always accurate. But very naiive and easy to implement.
- Option 2: Mean of recent past: For example taking mean of only past year data. This is going to be more accurate than overall historical mean since any recent trend will be captured in this mean.
- Option 3: Last Value: Simply taking the last observed value at t=n and considering it as predited value of t=n+1. However, this also does not address the concern of seasonality. For example, we might take mean from Q4 of last year as the predicted earnings profit for Q1 of this year, but this could be guided my quartely seasonality in the data.
- Option 4: Seasonal Mean: Taking mean of data for Q1 from last year to predict the expected value of earnings for Q1 of this year, and similarly extending the same concept for other quarters. In the naiive approaches this is the best approach.
- Naiive Error Calculations: Use MAPE which is Mean Average Percentage Error. This is computed by find the percentage deviation of each predicted data points from observed datapoints; and taking absolute mean of all these percentage errors. In computing each individual percentage error, the denominator is actual or observed value.