Search

We propose a limit order book (LOB) model with dynamics that account for both the impact of the most recent order and the shape of the LOB. We present an empirical analysis showing that the type of the last order significantly alters the submission rate of immediate future orders, even after accounting for the state of the LOB. To model these effects jointly we introduce a discrete Markov chain model. Then on these improved LOB dynamics, we find the policy for optimal order choice and placement in the share purchasing problem by framing it as a Markov decision process. The optimal policy derived numerically uses limit orders, cancellations and market orders. It looks to exploit the state of the LOB summarized by the volume at the bid/ask and the type of the most recent order to obtain the best execution price, avoiding non-execution and adverse selection risk simultaneously. Market orders are used aggressively when the mid-price is expected to move adversely. Limit orders are placed under favorable LOB conditions and canceled when non-execution or adverse selection probability is high. Using ultra high-frequency data from the NASDAQ stock exchange we compare our optimal policy with other submission strategies that use a subset of all available order types and show that ours significantly outperforms them.

This research attempts to model the stochastic process of trades in a limit order book market as a marked point process. We propose a semi-parametric model for the conditional distribution given the past, attempting to capture the effect of the recent past in a nonparametric way and the effect of the more distant past using a parametric time series model. Our framework provides more flexibility than the most commonly used family of models, known as Autoregressive Conditional Duration (ACD), in terms of the shape of the density of durations and in the form of dependence across time. We also propose an online learning algorithm for intraday trends that vary from day to day. This allows us both to do prediction of future trade times and to incorporate the effects of additional explanatory variables. In this paper, we show that the framework works better than the ACD family both in the sense of prediction log-likelihood and according to various diagnostic tests using data from the New York Stock Exchange. In general, the framework can be used both to estimate the intensity of a point process, and to estimate a the joint density of a time series.

We give an extension of de Finetti's concept of coherence to unbounded (but real-valued) random variables that allows for gambling in the presence of infinite previsions. We present a finitely additive extension of the Daniell integral to unbounded random variables that we believe has advantages over Lebesgue-style integrals in the finitely additive setting. We also give a general version of the Fundamental Theorem of Prevision to deal with conditional previsions and unbounded random variables., Comment: 21 pages

We derive generalization error bounds for traditional time-series forecasting models. Our results hold for many standard forecasting tools including autoregressive models, moving average models, and, more generally, linear state-space models. These non-asymptotic bounds need only weak assumptions on the data-generating process, yet allow forecasters to select among competing models and to guarantee, with high probability, that their chosen model will perform well. We motivate our techniques with and apply them to standard economic and financial forecasting tools---a GARCH model for predicting equity volatility and a dynamic stochastic general equilibrium model (DSGE), the standard tool in macroeconomic forecasting. We demonstrate in particular how our techniques can aid forecasters and policy makers in choosing models which behave well under uncertainty and mis-specification., Comment: 34 pages, 3 figures

Vapnik-Chervonenkis (VC) dimension is a fundamental measure of the generalization capacity of learning algorithms. However, apart from a few special cases, it is hard or impossible to calculate analytically. Vapnik et al. [10] proposed a technique for estimating the VC dimension empirically. While their approach behaves well in simulations, it could not be used to bound the generalization risk of classifiers, because there were no bounds for the estimation error of the VC dimension itself. We rectify this omission, providing high probability concentration results for the proposed estimator and deriving corresponding generalization bounds., Comment: 11 pages

The literature on statistical learning for time series often assumes asymptotic independence or "mixing" of the data-generating process. These mixing assumptions are never tested, nor are there methods for estimating mixing coefficients from data. Additionally, for many common classes of processes (Markov processes, ARMA processes, etc.) general functional forms for various mixing rates are known, but not specific coefficients. We present the first estimator for beta-mixing coefficients based on a single stationary sample path and show that it is risk consistent. Since mixing rates depend on infinite-dimensional dependence, we use a Markov approximation based on only a finite memory length $d$. We present convergence rates for the Markov approximation and show that as $d\rightarrow\infty$, the Markov approximation converges to the true mixing coefficient. Our estimator is constructed using $d$-dimensional histogram density estimates. Allowing asymptotics in the bandwidth as well as the dimension, we prove $L^1$ concentration for the histogram as an intermediate step. Simulations wherein the mixing rates are calculable and a real-data example demonstrate our methodology., Comment: 30 pages, 8 figures. Longer version of arXiv:1103.0941 [stat.ML]

We derive generalization error bounds for stationary univariate autoregressive (AR) models. We show that imposing stationarity is enough to control the Gaussian complexity without further regularization. This lets us use structural risk minimization for model selection. We demonstrate our methods by predicting interest rate movements., Comment: 10 pages, 3 figures. CMU Statistics Technical Report

The literature on statistical learning for time series assumes the asymptotic independence or ``mixing' of the data-generating process. These mixing assumptions are never tested, nor are there methods for estimating mixing rates from data. We give an estimator for the $\beta$-mixing rate based on a single stationary sample path and show it is $L_1$-risk consistent., Comment: 9 pages, accepted by AIStats. CMU Statistics Technical Report

Correction to Annals of Probability 29 (2001) 1612--1624 [doi:10.1214/aop/1015345764]., Comment: Published at http://dx.doi.org/10.1214/009117905000000503 in the Annals of Probability (http://www.imstat.org/aop/) by the Institute of Mathematical Statistics (http://www.imstat.org)