LessonThis is a lesson restated in new words. Any figure in it is an example from the source, not this desk's record.
A model can be the right family and still produce absurd probabilities if the parameters are stale. For the VIX, the lecture contrasts a mean-reverting specification, which stays in a plausible range, with a straight line that runs off to infinity after a spike. Inside the mean-reverting family, one parameter set says a VIX above 60 is a once-in-65-billion-years event. Another says about once every 4.2 years. A Kalman filter does not replace that model. It blends the model's one-step guess with the latest quote, and a single noise parameter decides which one it trusts.
Why it matters
Without the true dynamics you have to specify a model M and a parameter vector theta. The lecture's examples are a regression, an AR(1), a GARCH, and an Ornstein-Uhlenbeck mean-reversion model. A linear fit is a fair specification when the scatter is linear, and a poor one when the scatter bends, unless features are transformed. The linear model is linear in its parameters, not necessarily in the raw inputs. Volatility has a large literature as a mean-reverting process. Ignore that and a forecast after a spike extrapolates without a pull back. Respect it and the path oscillates around a long-run level that can itself change.
Even a theoretically motivated model is only as sane as theta. Method of moments and maximum likelihood have classroom properties the lecture does not treat as permanent. A gap down can be a new level and a new uptrend. A regression that keeps every old point still predicts a downtrend. A regression that adapts to the new regime does not. You will need a new parameter set, a new model, or both.
The idea in plain words
The filter's target is a latent state, here the VIX level, observed with noise. For the mean-reversion example the lecture names four pieces: F, the decay or speed; B, the long-run mean; Q, the process-noise variance; R, the measurement-noise variance. The Kalman gain is the model's uncertainty divided by that uncertainty plus the measurement uncertainty. Large R means a small gain and more trust in the model. Small R means a large gain and more trust in the quote.
How to use it
The worked VIX build has three steps. Offline, discretize an Ornstein-Uhlenbeck model as an AR(1), regress to get the AR coefficients, and back out the speed and the long-run mean. The window is a judgment. The whole history is right only if one distribution produced it. A picture in the lecture shows a shorter window matching the current regime better. Initialize the state at the first quote or at the long-run mean, with a large initial variance. After a burn-in, the gain settles.
Each new quote then forecasts one step, computes the innovation, the surprise relative to the mean-reversion guess, forms the gain, and updates the state and the variance. On real VIX data the filter tracks the series and the multi-step marks lean toward the long-run mean, including after a spike, where the forecast pulls back down. Illiquid bonds are the other application. A model price from Treasuries plus a credit spread is the forecast. Rare prints, including prints on similar issues, are the measurements. The lecture points at Bloomberg's BVAL function as this idea in production.
What usually goes wrong
A linear forecast of a mean-reverting series invents levels, including negative ones, that the subject does not visit. A mean-reverting model with a tiny probability of a VIX above 60 throws away a state the 4.2-year parameterization says you should plan for. Those errors are upstream of any trade.
Questions people ask
What is the 65-billion-year figure?
It is the badly parameterized mean-reversion model's implied wait for a VIX reading above 60. A better parameter set in the same lecture puts that wait near 4.2 years.
What does a large measurement noise do?
It shrinks the Kalman gain, so the filter trusts the model and moves slowly when quotes jump. A small measurement noise does the opposite.
Does the filter re-estimate the mean-reversion speed?
Not in the basic version. The speed and the long-run mean are fit offline. The filter updates the level. A dual filter is the extension that also updates parameters.