Forecastability

Generally, there are four conditions that determine forecastability.

_{Icons sourced from Getty Images.}

For example, for short-term forecasts of residential water use, the first three are usually satisfied:

1. There is ample data on historical water use and the factors that contribute to it.
2. We understand many of these contributing factors – such as house size, season, location.
3. In the short term (a few weeks), we can assume that behaviour will be similar to what has been seen in the past.

The final condition is less certain, because forecasts about water use can affect behaviour due to environmental and economic concerns.

Nevertheless, we could easily make a model that links residential water use and the key driver variables. As long as this model’s forecasts aren’t seen by residential water users, it would not affect their water use.

Share your thoughts

The following graph is based on arrivals in Australia from the UK between 1981 and 2012. The larger portion on the left of the graph visualises this historical data. The smaller, shaded portion on the right visualises forecasted arrivals for three more years. This forecast was made by a triple exponential smoothing algorithm within BigML.

_{Based on information from Australian Bureau of Statistics and Hyndman, R. J., & Athanasopoulos, G.}

For this course, you don’t need to know about what triple exponential smoothing is. Instead, consider the forecast in relation to the conditions above. How well do you think it meets them? Which conditions would it satisfy easily, and which ones are harder? Share your thoughts in the comments.

References

_{Australian Bureau of Statistics. Overseas Arrivals and Departures, Australia. ABS.}

_{Hyndman, R. J., & Athanasopoulos, G. (2021, May). Forecasting: principles and practice. OTexts.}