US 12,373,669 B2
Method and system for using deep video prediction for economic forecasting
Zhen Zeng, Ypsilanti, MI (US); Tucker Richard Balch, Suwanee, GA (US); and Maria Manuela Veloso, New York, NY (US)
Assigned to JPMORGAN CHASE BANK, N.A., New York, NY (US)
Filed by JPMorgan Chase Bank, N.A., New York, NY (US)
Filed on Oct. 28, 2021, as Appl. No. 17/452,676.
Claims priority of provisional application 63/107,877, filed on Oct. 30, 2020.
Prior Publication US 2022/0138778 A1, May 5, 2022
Int. Cl. G06Q 40/04 (2012.01); G06N 3/045 (2023.01)
CPC G06N 3/045 (2023.01) [G06Q 40/04 (2013.01)] 17 Claims
OG exemplary drawing
 
1. A method for forecasting a change in a market, the method being implemented by at least one processor, the method comprising:
receiving, by the at least one processor, historical market data;
converting, by the at least one processor via a video prediction network that uses computer vision techniques and includes a neural network, the historical market data into a plurality of first image video frames that correspond to a predetermined time sequence, wherein the historical market data includes a respective plurality of daily market closing values for each of a predetermined plurality of financial assets, and wherein the plurality of first image video frames includes, for each respective financial asset from the predetermined plurality of financial assets, a corresponding rectangle that is color-coded based on each respective percentage change in value over a predetermined time period, such that the percentage of change is converted into a corresponding pixel value representing the corresponding color-code of the corresponding rectangle;
inputting, by the at least one processor via the video prediction network, the plurality of first image video frames into the neural network to train the neural network for forecasting future images;
generating, by the at least one processor via the video prediction network, at least one future image video frame that corresponds to a future time point with respect to the predetermined time sequence, wherein the video prediction network applies the computer vision techniques and the neural network to the plurality of first image video frames to generate the at least one future image video frame; and
determining, by the at least one processor via the video prediction network, a prediction of future market data based on the at least one future image video frame,
wherein the generating of the at least one future image video frame further includes:
encoding, via a content variable, static content of the plurality of first image video frames;
inputting at least one latent state variable into a multilayer perception (MLP) to learn first order movement of the at least one latent state variable;
encoding, via a CNN-based encoder network, the plurality of first image video frames into a plurality of first image encoded video frames;
inferring an initial latent state from at least one of the plurality of first image encoded video frames;
inferring, via a long short-term memory, latent dynamics from the plurality of first image encoded video frames; and
concatenating, via a decoder network, the content variable and the at least one latent state variable for generating of the at least one future image video frame.