The following are my learnings from the Meetup talk by Prof Hyndman.

  • The following are the relevant timeseries based packages in R
    • forecast
    • tseries
    • Mcomp
    • fma
    • expsmooth
    • fpp2
  • Completely automated forecasting came out of the need to generate forecasts for thousands of series every week
  • Surprisingly good forecast is Naive for stock markets
  • Drift method - Random walk + average of first and last observation
  • croston : used in supply chain forecasting when the demand is intermittent
  • forecast class
    • original series, point forecast, prediction interval, forecasting method, residuals and in-sample one step forecasts
  • MAPE is not useful when the actual value is close to 0 as MAPE shoots
    • Also relevant for only if the target variable has a natural 0 and the target variable is always greater than equal to 0
  • MASE is better as it scale variant - stable measure of scale. the mean absolute error is scaled by mean one step forecast error for the training sample

\begin{equation} C = W\log_{2} (1+\mathrm{SNR}) \end{equation}

\begin{equation} MASE = n^{-1} \sum^n_{t=1} { |y_t-f_t| \over q1 } \end{equation}

  • 10 years of effort to develop the statistical properties of exponential methods
    • All the math is covered in “Forecasting with Exponential Smoothing”
  • 5 Types of trend, 3 Types of Seasonality - 15 Models with Additional error
    • Trend
      • None
      • Additive
      • Damped Additive
      • Multiplicative
      • Damped Multiplicative
    • Seasonality
      • None
      • Additive
      • Multiplicative
  • 5 Types of trend, 3 Types of Seasonality - 15 Models with Multiplicative error
    • Trend
      • None
      • Additive
      • Damped Additive
      • Multiplicative
      • Damped Multiplicative
    • Seasonality
      • None
      • Additive
      • Multiplicative
  • ETS function can fit 30 types of models - (Error, Trend, Seasonal)
    • 11 of them are numerically unstable
    • 19 of them are the ones which algo checks out
  • Damped Trend
    • Starts off with a trend and then goes down
  • Auto ETS works better than any other method for first 4 periods for all seasonal time series
  • Why use automated forecasts?
    • Most experts can’t beat automated forecasts
    • Not everyone is an expert
    • Many industries needed thousands of forecasts
  • Box Cox to stabilize variance
  • Animation of a set of graphs is possible in beamer
  • fpp2 builds on arima from the base package
  • ets() can only fit maximum period of 24
  • Arima() has a max period of 350 but runs out of memory if period >200
  • STL approach works pretty well. It allows decomposition of any frequency
  • dshw allows two seasonal periods
  • forecast function takes several types of timeseries objects