- Accuracy of naive forecasts is often difficult to beat by more complex models
- Incorporating naive forecasts is a way to acknowledge uncertainty
- Naive forecasts do not correlate with other forecasts, which makes them an ideal addition to a combined forecast
A Case for Naive Models
In forecasting, complexity often harms accuracy.
Very basic models, like the naive no-change model, can be challenging to surpass with more complex counterparts. The decision to employ a naive model such as the no-change model is often driven by the anticipation that the situation will remain stable or the inability to predict the direction of change.
In other words, incorporating a naive component is a deliberate strategy to recognize the underlying uncertainty in a situation and, in doing so, align with the principle of conservatism, as the Golden Rule of Forecasting.
Naive forecasts of the popular vote
In 2020, the PollyVote embraced the principle of conservatism by introducing naive forecasts as a new component. This addition, consisting of the average of forecasts from two naive models – the electoral cycle and a 50/50 model – further improved the accuracy of the PollyVote. This improvement was evident in both providing forecasts ahead of time (ex ante) and retrospectively for previous elections (ex post).
The Electoral cycle model proposed by Helmut Norpoth predicts election outcomes based on historical electoral patterns. The model uses the vote of the two most recent elections as predictor variables in a linear model, estimated based on data from all elections since 1828. The model’s vote equation reads as
V = 50.4 + 0.46 Vt-1 – 0.49 Vt-2,
where V is the Democratic party’s popular two-party vote, Vt-1 is the Democratic party’s popular two-party vote in the last election, and Vt-2 is the Democratic party’s popular two-party vote in the second to last election.
The 50/50 model is a simple, naive model, which assumes that both major-party candidates will gain 50% of the popular vote. The model thus represents the age of political polarization.