State Space Statistical Arbitrage
Building upon the ML focus from Trading the Unobservable, Triantafyllopoulos and Montana recently published a state space model for statarb spreads entitled Dynamic Modeling of Mean-reverting Spreads for Statistical Arbitrage. This article builds upon the 2005 Elliott et al. article Pairs Trading in JQF.
Contribution of this paper, per the abstract:
Building on previous work, we embrace the state-space framework for modeling spread processes and extend this methodology along three different directions. First, we introduce timedependency in the model parameters, which allows for quick adaptation to changes in the data generating process. Second, we provide an on-line estimation algorithm that can be constantly run in real-time. Being computationally fast, the algorithm is particularly suitable for building aggressive trading strategies based on high-frequency data and may be used as a monitoring device for mean-reversion. Finally, our framework naturally provides informative uncertainty measures of all the estimated parameters.
For those who have traded statarb using PCA or cointegration techniques (or are familiar with Statistical Arbitrage in the U.S. Equities Market by Avellaneda and Lee), this article provides interesting theoretical comparison.
Quantivity 
Interesting article, but from my point of view a bit too theoretical. It still feels like proposing a model (be it autoregressive or other) and then fitting it to the observation data. Given the ever changing nature of stock prices I am not sure if any model can be sustained for more than a brief period of time. My personal preference goes to model-less strategies, based on probability mappings.
@sjev: thanks for your comments. From a cursory look at your most recent post, looks interesting. Look forward to reading more in-depth.
I’ve read this paper recently and agree with sjev. What they do in the paper is basically manipulate the spread to represent it in state-space form as a kind of unobserved component + noise model. Although they show estimation results, they fail to show backtesting of any kind of trading with this model.
@Dr. Nickel: agreed. Coincidentally, I am finishing up a post on overfitting and data snooping pathologies, which seeks to cover a bit on these sorts of methodological shortcomings.