If the data points are equally spaced, the Henderson filter and the second order LOESS filter of the same length effectively deploy a very very similar set of symmetric weights. They essentially do the same job across the middle of the scatterplot. If the data points are not equally spaced (and with opinion polls they are not), then the LOESS regression technique is technically better.

However, their treatment of the end point problem is a little different. Both effectively use a system of asymmetric weights. Arguably, the LOESS system has a higher weight on the final data points than the Henderson system. While this makes it more responsive to identifying turning points, it also leads to “false positives” and it more readily mistakes noise at the end of a series as signal. (The converse argument is that Henderson is a little more laggy when it comes to the turning points). But quite honestly, that is a quibble.

I use Henderson, primarily because it is simpler with my software. It looks like Possum uses something like the R Stats package, and so I suspect he uses LOESS for the same reason.

But seriously, both approaches serve the same purpose. They remove the noise from the polling data to reveal the underlying signal.

There’s nothing wrong with HMA’s – they’re one of many types of smoothing algorithm. I just prefer LOESS regressions because of their flexibility.

I just use Gallup because it’s a large poll, with a rolling three day average (meaning it has relatively low levels of volatility) – what’s important about it as a poll isn’t it’s headline results per se, but any slow trends that emerge from it. If Gallup is, say, showing a trend toward Obama and the markets are doing the same, then we know that movement is consistent and likely to be real.

If the markets stay at a consistent value of Electoral College votes, and the Gallup stays consistent, then it’s pretty likely that the contest isn’t changing from where it recently was. We could use any of the Daily tracking polls for this, like Rasmussen for example – I just had to choose one, so Gallup it was.

This seems to be another method of smoothing out the noise – I wonder if you would care to give your thoughts on how it compares with the LOESS regression?