Generative network based probabilistic portfolio management

A computer program product comprising a computer readable storage medium having program instructions embodied therewith, the program instructions executable by a device to cause the device to:

• create a training data set including at least asset-price trends of a plurality of assets over a past time period and a random latent vector sampled from a prior distribution associated with the asset-price trends of a plurality of assets, wherein the training data set includes a time series data, the asset-price trends being normalized, compressed and concatenated with an analysis vector comprising an analysis value associated with an asset for each of the plurality assets, the analysis value computed as a ratio of, a difference between a maximum price and a minimum price, and a mean price associated with the past time period;
• train a deep-learning neural network to model a probability distribution of the asset-price trends for a future time period using the training data set, wherein the deep-learning neural network receives the time series data as input, the deep-learning neural network automatically learning non-linear dependencies between the plurality of assets, the deep-learning neural network architected to include a first neural network being trained to generate simulated future prices of the plurality of assets and a second neural network being trained to discriminate the simulated future prices from real prices, the first neural network and the second neural network being trained in an adversarial training process, wherein the second neural network is trained to minimize an output value for the real prices and maximize the output value for simulated future prices, while the first neural network is trained to generate the simulated future prices that minimize the second neural network's output value;
• execute a portfolio optimization on the modeled probability distribution generated by the trained deep-learning neural network;
• and
• estimate expected risks and returns for different portfolio diversification options based on the portfolio optimization,
• wherein the first neural network includes at least consecutive layers that compress an inner representation of the training data set and a sequence of transpose layers that expand the time resolution, wherein at every layer of the transpose layers the time resolution is doubled.

A system comprising:

• a hardware processor;
• a memory device coupled with the hardware processor;
• the hardware processor operable to at least:
• create a training data set including at least asset-price trends of a plurality of assets over a past time period and a random latent vector sampled from a prior distribution associated with the asset-price trends of a plurality of assets, wherein the training data set includes a time series data, the asset-price trends being normalized, compressed and concatenated with an analysis vector comprising an analysis value associated with an asset for each of the plurality assets, the analysis value computed as a ratio of, a difference between a maximum price and a minimum price, and a mean price associated with the past time period;
• train a deep-learning neural network to model a probability distribution of the asset-price trends for a future time period using the training data set, wherein the deep-learning neural network receives the time series data as input, the deep-learning neural network automatically learning non-linear dependencies between the plurality of assets, the deep-learning neural network architected to include a first neural network being trained to generate simulated future prices of the plurality of assets and a second neural network being trained to discriminate the simulated future prices from real prices, the first neural network and the second neural network being trained in an adversarial training process, wherein the second neural network is trained to minimize output value for the real prices and maximize the output value for simulated future prices, while the first neural network is trained to generate the simulated future prices that minimize the second neural network's output value;
• execute a portfolio optimization on the modeled probability distribution generated by the trained deep-learning neural network;
• and
• estimate expected risks and returns for different portfolio diversification options based on the portfolio optimization,
• wherein the first neural network includes at least consecutive layers that compress an inner representation of the training data set and a sequence of transpose layers that expand the time resolution, wherein at every layer of the transpose layers the time resolution is doubled.