If you’re not familiar with TensorFlow, a good place to start is with the self-study introduction notebooks listed here. Next, see the examples we offer here demonstrating usage of TFF methods. You can use Google Colab to run these examples on GPUs for free.

Avera AI is part of Area 120 by Google, which is a workshop for Google's experimental products, helping small teams rapidly build new products in an entrepreneurial environment.

No, the TensorFlow side of the service can be used and tested for any part of the portfolio. All components are easily accessible an can be run independently if needed. For example, low level TensorFlow Quant Finance tools can be used for American Option pricing under the Black-Scholes model. See it here.

Integration effort will vary depending on internal solutions of the customer. For example, if single codebases and data repositories exist. We will be partnering with third party consulting agencies to advise on the integration effort on a case-by-case basis and also provide support throughout the integration process.

Customers can engage in multiple ways:

• Use TF Quant Finance Library and execute on-prem
• Use TF Quant Finance Library and execute through the Pricing Engine; TFQF can work on any machine, but is most efficiently executed through the Pricing Engine
• Plug into the Risk Reporting Engine by providing sensitivities
• Use the service as an end-to-end solution

The roadmap includes: ODE solving framework; Advanced generic SDE sampling schemes (e.g., Milstein, Runge-Kutta); Multidimensional interpolation schemes; Continuously expanding; Model coverage (Local volatility, SABR); Model calibration tools; Local stochastic volatility calibration framework; Rates and FX Instrument module coverage.

Not seeing your request in the roadmap? Please file a bug.

Pricing American put options under Black-Scholes model using a PDE:

• Quantlib: 125 options/second
• TensorFlow Quant Finance on a CPU: 308 options/second
• TensorFlow Quant Finance on a GPU: 5531 options/second

Pricing American put options under Heston using a PDE (100 timesteps, 500 x 500 spatial grid):

• Quantlib: 0.15 options/second
• TensorFlow Quant Finance on a CPU: 0.6 options/second
• TensorFlow Quant Finance on a GPU: 6.25 options/second

With the implementation of FRTB, as indicated in Basel III and Basel IV regulations, imminent, and considering the immense computational burden this will place on banks, we thought FRTB was where our initial efforts would be best spent. Since all banks will have to comply using the Standardised Approach, and a subset may qualify for the Internal Models Approach - this is the order in which we’ve prioritised these regulations.

We plan to incorporate additional regulations in our Risk Reporting Engine’s coverage according to priorities from clients and the industry.

TF in itself is not AI, but can speed up pricing valuations and risk factors significantly through automatic differentiation and removal of redundant computations. Using AI will make this even more efficient by optimally allocating computing resources. Other AI functionalities can reduce errors (through Anomaly Detection) and increase explainability (through Fair Attribution of capital to specific trades or portfolios).

No, we will be providing integration with Anthos which will allow our services to run on other cloud platforms.