chelynl / esm_time_series

Forecasting monthly ozone levels with Exponential Smoothing Models in R

Ozone Air Quality: Project Overview

We want to explore the average monthly concentration of ozone (ppm) in a local area.

Part 1: Exploring the Data

• The data recorded daily max 8-hour ozone concentration per observation.
• Looked for missing values and noted which year had the most missing values
• Rolled up data to the monthly level by using the average of the observations present
• Created a time plot of the mean monthly max 8 hour ozone concentration
• Evaluated time plot for potential trends or seasonality captured

Part 2: ESM Model Building

• Split the data into training (withholding last 17 months), validation (next 12 months), test (last 5 months) data sets
• Created a monthly ESM forecast with the training data set
• Made data visualizations of the following:
  • Actual Ozone values overlaid with the trend/cycle component for the training set
  • Actual Ozone values overlaid with the seasonally adjusted Ozone values for the training set
  • Time Plot of the predicted versus actual for the validation and test data
• Decomposed time series into respective componenets (raw data, trend, seasonality, noise) using STL
• Explored additive ESM vs. multiplicative ESM
• Calculated MAPE and MAE values for evaluating accuracy of forecasts with test data

Part 3: Testing Stationarity and White Noise

• Checked for stationarity of the average monthly ozone levels including any potential trend and/or random walks
  • used the Augmented Dickey-Fuller tests up to lag 2 tests for the results
• Implemented techniques for getting stationary data such as taking the first difference
• Tested whether the stationary series exhibited white noise using the Ljung-Box test
• Explained implications for futher ARIMA modeling

Data Visualizations: