Beyond the decimal point
Sesame Q is a quantum finance laboratory. We combine machine learning with quantum and quantum-inspired algorithms to value loans and to decide which loans to hold, starting with real estate credit.
The average future never arrives.
A loan against a home
The average loan does not default. A loan against a home repays in full in most futures. In a few, prices fall, the sale takes a year and the proceeds come up short. Those few decide what the loan is worth. We value the home as a range, the loan in each of its futures, and the terms that follow.
The same model extends from one loan to a book and the capital behind it, to lending against vehicles, equipment and commercial property, and to loans that are bought, sold and insured.
The rate paid to savers
The average saver does not leave. When a rate changes, most balances stay. Some go, and they can go in a day, while the loans they funded repay over years. We value balances together with the right to withdraw them, and from that set the rate on each account and the cash held against a bad month.
The same model extends to banks and credit unions setting savings and certificate rates, to brokerages and insurers crediting cash and annuities, and to card and lending programs pricing rewards, fees and limits.
Open problems
Each of these is an open problem we intend to explore. We work on them parametrically: a book is described by a short list of fitted parameters, such as default, prepayment and withdrawal rates, rate volatility and the correlations between them, and the model is re-run as those parameters move.
Valuation
- Loan valueThe value of a loan at origination across rate, prepayment and default scenarios, with a confidence interval.
- Rate pricingThe rate each loan should carry for its term, leverage and risk, and the value of the interest-rate caps on floating-rate loans.
- CollateralA home’s value as a distribution, and its neighborhood’s, updated as sales, rents and rates change.
- Savings balancesWhat a balance is worth to the institution holding it, when the saver may withdraw at any time.
- Tail risk and capitalThe loss in the worst one in two hundred outcomes, and the capital it implies.
- Forward valueThe value of a book at a future date, in each of the futures that lead there.
- Securitized creditBonds backed by mortgages and loans, valued through the rules that divide their cash flows.
- Servicing rightsThe value of collecting a loan’s payments, which ends when the borrower prepays or defaults.
- Insurance guaranteesLong-dated promises whose cost depends on markets and on how policyholders behave.
Optimization
- Reward ratesThe rate paid on each account and term, set against how balances respond when it moves.
- Funding and liquidityThe cash to hold and the funding to use, so that lending continues when withdrawals rise and repayments slow.
- Loan termsThe size, term and extension options that make a loan worth making.
- Portfolio constructionWhich loans to fund, hold and sell among thousands, within limits on concentration, liquidity and capital.
- Funding linesWhich loans to pledge to which facility, when each lender sets its own rules.
- HedgingThe lowest-cost set of rate and prepayment hedges for a whole book.
- WorkoutsFor each loan near maturity, whether to extend, modify, sell or foreclose, chosen across the whole book within loss and capital limits.
- Capital allocationThe mix of lending that earns the most for the capital it ties up.
- RebalancingMoving a portfolio to its target in whole lots, net of the cost of trading.
How the laboratory works
Machine learning supplies the inputs. Quantum and quantum-inspired algorithms do the valuation and the optimization.
- 1
Learn
Machine learning turns data lenders already hold into a short list of parameters: default, prepayment and withdrawal rates, and value paths for each property.
- 2
Precompute
The hard cases are solved ahead of time, across the range those parameters can take.
- 3
Map
Software places today’s case among the solved ones. It interpolates a value, or adapts a solved decision.
- 4
Decide
The answer arrives in time to act, with its interval.
On quantum hardware
The second step is the one that moves. On fault-tolerant hardware, amplitude estimation values a book to an error ε with on the order of 1/ε samples, against 1/ε² for Monte Carlo: about ten thousand against a hundred million at one part in ten thousand. For optimization, parameterized circuits search among portfolios. Both depend on loading the inputs cheaply, which is why our models are parametric.
What runs today
The models run now on classical and quantum-inspired methods. Each is tested on the same loans as the lender’s current method, and the comparison is reported with the result. On today’s quantum hardware neither problem runs faster than a good classical method. We will claim a speed-up when we have measured one.
Founders
T Zhu and Robbie Yan have built consumer financial products, lending and risk systems for financial institutions, and trading models run in live markets.
T Zhu
Co-founder and Chief Executive
Two decades building companies. T co‑founded Tellus, a financial technology company spanning savings, real estate lending and property management. Before that he led Sodacan, a mobile company, as chief executive, and was chief technology officer of More Corn, acquired by Answers.com. Stanford.
Robbie Yan
Co-founder, Chief Scientist and Chief Technology Officer
Two decades building artificial intelligence and quantitative systems for finance. Robbie co‑founded Advance Intelligence Group, whose risk, fraud and lending systems serve financial institutions across Asia, and Shanghai Jiwei, a quantitative investment firm trading Asian equities. Earlier he worked on machine learning at Google, Citadel and Two Sigma. PhD in electrical engineering, Stanford.
Join the laboratory
We are recruiting a Head of Quantum Algorithms, and researchers and engineers in quantum algorithms, machine learning, quantitative finance and financial software. Write with a short note on your work and one paper or repository we should read.
Correspondence
For the investor letter, design partnerships, or to join the laboratory.