foresight machines
Introducing Foresight MachinesWe’re building AI that models financial and physical systems, projects how they could unfold, and searches for better ways to allocate capital and coordinate resources.
Forecasting anticipates the conditions ahead, and machines now do it about as well as the best people.
But every commitment sits inside a system larger than the decision itself: other participants, agreements, constraints, and the conditions they meet. The outcome turns on how they move together.
Foresight runs that system forward, then searches for a better arrangement: a different allocation, new terms, a sequence that makes a project possible.
Models like these took specialists months to build. AI is making them fast to build and practical to keep current.
Foresight Machines is building AI to expand the choices institutions can make, and the futures those choices make possible.
01Dashboards
Show what happened.
You read them.
02Forecasts
Estimate what is likely.
You interpret them.
03Foresight Machines
Project what could happen and find what could work.
You set the goal and decide.
Ask what could happen, compare what you could do, or start from the outcome you need.
01Conditions
Multi-brand digital-dollar issuer
Plan
Reserves sit across four banks, none above the concentration limit.
Simulation
Two of the four route payments through the same correspondent. Lose it, and more than half the cash freezes at once.
Finding
The concentration is in the plumbing, not the balances.
02Choices
Bank treasury
Plan
Grow the buffer by a fifth.
Simulation
In the stresses that matter, the buffer is large enough but slow to turn into cash. A committed line closes the gap at lower carrying cost.
Finding
A smaller buffer, plus a committed line whose draw is tested regularly.
03Arrangements
Digital-dollar issuer
Plan
Same-day redemption only below a fixed size.
Simulation
Large redemptions cluster in stressed markets, exactly when Treasury bills are slowest to turn into cash.
Finding
A standing repo agreement that turns bills into cash within the day, sized to the largest cluster tested.
RequiresA repo counterparty and an operating test.
01Conditions
Private credit fund
Plan
Borrowers refinance as rates fall.
Simulation
Without refinancing, covenant breaches begin in one sector within eighteen months.
Finding
The risk is one sector’s refinancing wall, not the portfolio average.
02Choices
Asset manager
Plan
Sell bonds when cash runs short.
Simulation
In a rate shock, calls peak just as bond prices fall, so forced sales lock in the losses.
Finding
Borrow against the bonds through repo, and sell only once the calls subside.
03Arrangements
Equipment lender
Plan
Decline fleet operators with less than two years of history.
Simulation
Losses track the equipment’s resale value far more than the borrower’s age.
Finding
Lend against the equipment with a manufacturer buy-back, and raise the advance rate as payment history builds.
RequiresThe manufacturer’s buy-back agreement.
01Conditions
Regional property insurer
Plan
Reinsurance covers the largest single event.
Simulation
The second storm lands after the limit is spent, just as the renewal is priced.
Finding
The exposure is the second event, not the larger one.
02Choices
Regional property insurer
Plan
Raise the reinsurance limit.
Simulation
A reinstatement provision costs less, and covers the two-storm case the higher limit misses.
Finding
Keep the limit. Buy the reinstatement.
03Arrangements
Specialty insurer
Plan
Decline every home within a mile of the coast.
Simulation
Losses concentrate in homes without roof retrofits.
Finding
Offer cover on condition of a retrofit, with a premium credit for it.
RequiresAn inspection partner.
01Conditions
Solar-and-storage developer
Plan
Operations start before the first debt payment.
Simulation
Revenue slips past the first payment, and the storage contract’s delay penalty starts the same month.
Finding
The binding constraint is the penalty clause, not the loan.
02Choices
Electronics manufacturer
Plan
Double the safety stock.
Simulation
A second source for one controller chip covers most of the exposure, at a fraction of the carrying cost.
Finding
Qualify a second source for that one part.
03Arrangements
Infrastructure developer
Plan
One construction loan, sized for the full build.
Simulation
Debt coverage fails in the high-rate cases.
Finding
Draw the loan in stages with equipment deliveries, and pre-sell storage capacity to the utility.
RequiresThe utility’s agreement.
Composite examples, for illustration.
Every recommendation shows its sources, its assumptions, and how confident the model is.
01
Rules and relationships trace to operating records, contracts and observed behavior. Missing evidence is marked, not filled in.
02
Each recommendation is compared with current practice and strong conventional methods under the same objectives and constraints, including the conditions under which it fails.
03
The model is updated as contracts, counterparties and conditions change. Every version is archived, so any past recommendation can be reproduced.
04
Expectations are recorded before events unfold. Misses become proposed revisions, and a revision must improve results on separate cases before it is used.
On modeling, simulation and discovery.