UC San Diego

This thesis develops an efficient training method for optimizing path-dependent

performance measures in single asset trading scenarios using gradient descent.

We compute the exact, path-dependent gradients for both linear and nonlinear

decision models. We account for transaction costs as a fixed percentage

of any trading decision relative to the magnitude of the size in position

change. We apply our method to two measures of performance for price series of

T steps: mean return and the Sharpe ratio. For linear models, we achieve a

training complexity of O(TL) where L represents a fixed number of previous

time steps considered for the current decision. Neural networks with a single

hidden layer require O(TH^2L) to train for a hidden layer of size H. We

allow long positions, short positions, and decisions can scale to arbitrary

ranges to better fit real world scenarios like margin buying. Experimental

results on both synthetic and real-world price series demonstrate both correct

learning and the ability to capture and take advantage of mispricing in the

market via techniques like pairs trading.