At the end of the day though, any time you have time series data you can apply filters to smooth and shape your data. I don’t understanding how they’re modeling their data. There’s a good chance they’re doing some kind of frequency modeling where they’re counting correct predictions. It definitely sounds like they’re doing some stochastic modeling when they start talking about percentage predictions. You can definitely shape frequency domain as well with filters, though I havn’t quite thought through how the stochastic aspects might interact.
Keep in mind, filters are very basic, and even something as common as averaging data is a low-pass filter. As is fitting to a curve. This all acts to attenuate the signal we care about without also attenuating the noise. Though, again, if someone isn’t being rigorous about what constitutes noise, then no amount of filtering will actually help...
I’m also sorry if this thread isn’t very insightful. I’ve been having my nose rubbed in signal processing at work for the last 2 months, and it’s all I can see everywhere I look. I see parrallels everywhere that may not be there.
You’re also very correct about the simplicity of physical models versus social sciences. It may just be that trying too hard to apply basic information theory at models that are almost impossible to create in the first place is a fools errand.