Deep Learning in Financial Modeling

近四十年来，金融衍生品估值技术得到了长足的发展。然而，只有少数实际模型可以以封闭形式求解，并且由于大多数使用数值方法，如有限差分偏微分方程求解器，离散时间树或蒙特卡罗模拟器，这些传统方法相当缓慢。在数百个CPU内核上运行账面估值和风险流程需要一个通宵的过程，以符合监管和会计标准。这造成了巨大的成本（约1000万美元/年）和相应的环境和碳影响。由于会计准则IFRS 17，保险业现在也面临着类似的问题。此次合作的主要目标是开发深度学习，以及更普遍的机器和强化学习模型，以加速这些过程，并为资产价格、波动面、衍生品估值和对冲提供高效的模拟引擎。

The last four decades saw the development of the financial derivatives valuation technology. Only a few practical models, however, can be solved in closed form, and as most utilize numerical methods such as finite-difference partial differential equation solvers, discrete-time trees, or Monte-Carlo simulators, these traditional methods are quite slow. Running book valuation and risk processes on hundreds of CPU cores requires an overnight process to comply with regulatory and accounting standards. This creates a huge cost (~$10 million/year) and a commensurate environmental and carbon impact. Similar issues are now facing the Insurance industry due to the accounting standard IFRS 17.The key goal of the partnership is to develop deep learning, and more generally machine and reinforcement learning, models to accelerate these processes, and to provide efficient simulation engines for asset prices, volatility surfaces, derivative valuation, and hedging.