Time Series Prediction

Some details specific to time series:

• Feature engineering is difficult and should be handled by care (can easily learn bias - i.e. trends, seasonality, accumulation)
• Data is much smaller
• A different algorithm is needed
  • extrapolation
  • confidence interval
• Residuals should represent white noise
  • no correlation
  • zero mean and constant variance
• Resolution matters
  • prepare data and predict very granular then predictions can be aggregated, for example: predict use and predict hourly then take average for daily
• Confidence intervals:
  • keep some part of the data aka. dropout and run simulations
  • kernel mixture network
• Some models might have bias
  • have high accurary or high error (black or gray swan according to the nature of the series)

Models to be tested in the order of complexity (to have a baseline and to iterate):

• simple hand-made formula: average the data
• statistical model: autoregression such as SARIMA, linear sum of old observations
• exponential smoothing or winter-holt: this time exponential functions instead of linear, forget old observations at an exponentially decreasing reate
• simple neural network
• advanced neural network: LSTM, GRU, etc.