人工知能と金融工学,인공 지는

最近通貨先物を扱う自動注文管理と取引システムを開発しています。Three types of machine learning methods exist: supervised learning 教師あり学習, unsupervised learning　教師なし学習, and reinforcement learning　強化学習. The first type, supervised learning, includes techniques like linear least squares. It endeavors to find the set of weights which produces the least error in the hypothetical model when testing on training data for which the correct output is known. These weights are then used to estimate the output for new inputs. The second method involves classifying data based on patterns observed in it, without knowing the correct outputs of a training set. The last method involves providing a reward function to indicate the level of desirability for a given state and developing a policy of actions to take in any given state.
An example of using supervised learning in finance might be in a pricing model. At each time step, a vector of state information (volatility, underlying spot price, time to expiry, interest rate, etc) is given, and the system will compute whether the price of the future will rise or fall at the next time step. The training set will consist of pairs (state,movement), where movement is either "rise" or "fall" (amount of rise/fall is fixed at some small delta). By computing a linear combination of the input coordinates that produces the least error when run over the training set (i.e. rise=1, fall=0), a price movement prediction formula can be learned.
An example of unsupervised learning would be the observation of data points representing stocks and determining which points were correlated. For each stock, a vector consisting of price time series, volatility time series, and industry can be considered. Those stocks with a "distance" less than a threshold from the "center" of the group will be classified together. These stocks will be expected to have prices that move roughly in step. This type of classification can be used in statistical arbitrage, where stocks that are not in line with their group could be sold (if priced above the group) or bought (if priced below), while being hedged with a stock in the group (i.e. long the group stock if short the divergent stock).
An example of reinforcement learning might be in trading strategy selection. Reinforcement learning is often used in training robots to accomplish tasks like steering a vehicle. The usual model is a Markov Decision Process (MDP) defined by (S,A,D,P,R,gamma) where S is the set of states, A the set of actions, D the distribution of the initial state, P the transition probabilities, R the reward function, and gamma the discount factor. The goal of an algorithm is to compute the optimal policy, i.e. the action to take in each state maximizing the reward over the entire path. In trading, a state might consist of a portfolio, the values of the securities, interest rates, and the prices of other securities. Actions to be taken would be to buy/sell stocks within or outside of the portfolio to optimize return.