I’m working on developing a model for Portuguese elections, inspired by @AlexAndorra’s French Presidential model. You can find my current implementation here
As I’m new to this type of modeling, I’d greatly appreciate your thoughts on the following questions:
Gaussian Process (GP) Selection: With approximately 500 polls in total, would you recommend using the exact GP or HSGP? I’ve implemented HSGP because I couldn’t get the exact Latent to work with dimensions, though I notice this has been added recently. I’d appreciate any feedback on my current GP implementation.
Fundamentals Modeling: @AlexAndorra, in your French model, you included a component affecting vote share per party based on unemployment. Did this take into account whether parties were currently in government? I’ve been experimenting with using a switch to model the effects of party governments and economic sentiment. What are your thoughts on this approach?
Daily Snapshots and Posterior Updates: As the election approaches, I aim to provide daily snapshots of the race. However, each time I add new polls, I’ll recalculate the posterior from scratch, potentially changing previous days’ estimates. How would you recommend handling this to maintain consistency in reporting?
Thank you in advance for your insights and assistance!
How is convergence and sampling time looking right now? Also the size of the data. These are the main elements to look at to decide I’d say.
I did try it, but it didn’t add anything, so I got rid of it, to keep the model simpler. It’s very easy to implement though (just an indicator variable), so I’d encourage trying it and see if it really makes a difference.
I know it’s important in US elections for instance, but France’s moving party definitions apparently makes it harder for the voters to attribute responsibility – unless it’s a clear, big event, like Covid for instance.
I think you can just sample MCMC once, and then do out-of-sample posterior predictions anytime new polls come out – that’s if you don’t want to rerun MCMC each time.
Hi @bernardocaldas, interesting work thank you for tagging me. In my model I didn’t include fundamentals like unemployment, I just didn’t have enough historical data to say what effect they had and I assumed there would be a pass through into polling anyways. An example in this electoral cycle was that FiveThirtyEight and other models were relying on assuming a reversion to fundamentals, so showed Biden with a wide lead until June. When I made my model I was shocked to discover just how far behind Trump he was. The problem with assuming a reversion to fundamentals over time is that 1. you don’t know that the polls will converge to what you think they will and 2. there just isn’t evidence that there is such a convergence. It’s better to let the model show the uncertainty then to stack the deck for an economics or incumbency based result.
Andrew
Not to say fundamentals do not matter - I did some econometric analysis on Brexit results like a decade ago in undergrad and found that the concentration of immigrants from Eastern Europe was a factor in district-level results. But that kind of analysis is way easier to do post-facto.
