Skip to content
JoaFin
ES/EN

Approach

Mathematics first; AI to accelerate

There are two ways to get quantitative work wrong today: ignore AI and work ten times slower than necessary, or hand it your judgment and ship models nobody can defend. This is the middle we practice.

01

Formulation comes first

Most of the value is created before any code is written: deciding what is being modeled, what is assumed, what data actually exists, and by what criterion the model will be declared to work. A well-posed problem usually admits a simple solution; an ill-posed one is not fixed with more compute.

02

Methods are chosen for their properties

We choose a method for what can be proven about it — existence, uniqueness, stability, order of convergence, behavior under sparse data — not for its popularity. Sometimes that means a well-specified regression instead of a neural network, and sometimes exactly the opposite.

03

AI is a multiplier, not the author

We use language models to implement, refactor, review literature, generate tests and automate the repetitive parts. It multiplies execution speed. It does not transfer responsibility for the formulation, nor for verifying that the result is correct. Part of the doctoral research behind the firm is on the convergence of machine learning models under perturbation of their data: we know these tools from the inside, cracks included.

04

Nothing ships unverified

Every model is checked against analytic solutions where they exist, against independent methods where they do not, and against out-of-sample data always. If a result cannot be reproduced from scratch, it is not finished.

05

Assumptions are written down

Every assumption is documented with its justification, its source and its effect on the result. That is the difference between a model that survives staff turnover, an audit or a cross-examination, and one that does not.

06

You get something you can operate

The goal is not a report that gets filed, but a tool your team understands, runs and modifies. Knowledge transfer is part of the scope, not an add-on.

What AI does and does not do on our projects

We would rather be explicit, because the question matters and almost nobody answers it clearly. This is not a marketing stance: there is published mathematics, written in this house, on when a learning model converges and when it does not.

We do use it to

  • Implement and refactor code from a specification we write ourselves.
  • Review and synthesize technical and regulatory literature as a starting point.
  • Prototype several alternatives quickly to rule out the ones that fail.
  • Generate tests, edge cases and documentation.
  • Automate data acquisition, cleaning and repetitive tasks.
  • Act as an adversarial reviewer of our own reasoning.

We do not use it to

  • Decide the mathematical formulation of the problem.
  • Substitute for derivations, error bounds or convergence analysis.
  • Produce numbers that reach the deliverable without independent verification.
  • Justify assumptions: that requires domain knowledge and data.
  • Process confidential client information without prior, explicit agreement.
  • Write expert reports: the expert opinion is held by whoever signs it.
Fn(f)=Eπn[L(f(X),Y)] → Γ  F(f)=Eπ[L(f(X),Y)]F_n(f) = \mathbf{E}_{\pi_n}\bigl[ L(f(X), Y) \bigr] \ \xrightarrow{\ \Gamma\ } \ F(f) = \mathbf{E}_{\pi}\bigl[ L(f(X), Y) \bigr]

πₙ is the measure associated with the sample, π the true one, L the loss and f the fitted model. The Γ arrow denotes Γ-convergence; it is the right notion because it implies the minimizers converge, not merely the values.

Γ-convergence · Chapter 3 of the thesis

-2-1012θF(θ)
  • n = 5
  • n = 15
  • n = 60
  • F (true distribution)
F(θ) is the expected loss under the true distribution; Fₙ(θ) is the same loss under a sample of size n. At n = 5 the empirical minimum falls in the wrong basin entirely: the model confidently learned something that was not there. As n grows, Fₙ converges to F and the minimizers follow. Every curve is computed on one fixed sample.

Verification

What gets checked before a model leaves JoaFin.

  1. 01Dimensional and unit consistency throughout the computation.
  2. 02Comparison against closed-form solutions or known limiting cases.
  3. 03Convergence studies and sensitivity to discretization.
  4. 04Out-of-sample validation and backtesting by horizon.
  5. 05Sensitivity to each assumption, reported explicitly.
  6. 06Economic constraints: no-arbitrage, monotonicity, signs, conservation.
  7. 07Full reproducibility from raw data to final figure.
  8. 08Automated tests that fail when behavior changes.

What you receive when a project closes

  • The model, with its mathematical formulation written out.
  • Version-controlled, readable code with tests.
  • Documentation of assumptions, data and limitations.
  • Validation evidence and sensitivity results.
  • Instructions to reproduce everything from scratch.
  • A handover session with your team.

Have a problem that spreadsheets won't settle?

Get in touch