| dc.description.abstract | This thesis considers constructing high-frequency quantitative trading models. The work is a continuation of my project thesis (spring 2009) and Birgitte Ringstad Vartdal's master thesis (2000). We build our trading model through what we call the Layer Approach. This includes letting different layers take control of the different risk and decision mechanisms of our system. The underlying regression model is the Rydberg-Shephard model, the regression models are fitted to a moving data set to incorporate dynamics. We define a pool of model specifications, and introduce model selection tools, based on moving average payoff and scoring rules, to select the appropriate model specification. We introduce risk management tools for automatic switching the system on and off. Lastly, we apply a simple mean-variance technique to size our positions. We rank competing trading models through risk adjusted return. The theory underlying our approach is developed and discussed. Specifically, we develop an approximate test statistic to investigate the trade-off between bias and precision used to evaluate the moving data window size. The theory of scoring rules is discussed, particularly for categorical variables. We show that ranking models based on historical payoff is a proper scoring rule for two trading strategies. Four classical scoring rules (Brier, Spherical, Logarithmic and Zero-one) are introduced. We discuss the rationale behind the on/off-button, and derive the position sizing technique from a practical view point. Input parameters for the suggested trading model are estimated through extensive in-sample-testing (calibration). The calibrated input parameters are tested through 3 months of out-of-sample data for validation. The calibration data set is 1-minute data from August 2009 for the front month S&P500 futures contract. The out-of-sample data is 1 minute data from July, October and November (all 2009) for the same contract. The calibration indicates that the suggested trading strategy should only perform short selling (short only strategy). Our results show that historical performance leads future performance. Out-of-sample tests confirm the calibration results good. We conclude that both our suggested model selection tool and on/off-button probably increase the risk adjusted return. We find that our suggested short-only strategy is preferred for the out-of-sample data as well, and probably returns a positive return when not considering transaction costs. Comments regarding transaction costs are included. On the down side, we find that our simple position sizing technique does not perform as indicated in the calibration. | nb_NO |
