AlphaIQ mark

AlphaIQAutoSimAn AlphaIQ product

Describe any system in plain language. Get a validated simulation.

AutoSim reads your description or PDF, writes the simulator, fits it by Bayesian calibration and scores it out-of-sample — in your browser, in minutes. Every number traces back to an executed computation.

See the process. Change the outcome.

How it works

The problem

Serious decisions run on simulation. Building one is broken.

6 months
What a bank, insurer or supply-chain team spends building one production simulation model with a specialist quant team.
$500K+
Fully-loaded cost per model — before validation, documentation and model-risk review double it.
0 trust
What an LLM-generated model earns from a risk committee: no calibration, no out-of-sample test, no audit trail.

The new failure mode: anyone can ask an AI for a model that looks right. Nobody can tell you if its numbers mean anything. The bottleneck moved from writing models to trusting them.

The solution

The AI writes the model. Executed code decides if it is right.

The AI does

  • Reads your description or PDF
  • Proposes parameters with literature citations
  • Writes the Python simulator, streamed live
  • Diagnoses failures and proposes repairs

Executed code does

  • Runs real Python in your browser via WebAssembly
  • Fits parameters by Bayesian calibration (ABC-SMC)
  • Scores validity out-of-sample on held-out data
  • Exports an auditable model-risk report

AI narrative is never folded into a numeric score. Every number traces to an executed computation.

How it works

Six stages, one pipeline — description to audit-ready model.

Each stage produces an artifact you can read, and nothing downstream runs on a claim the stage before it could not support.

  1. 01

    Model

    Plain-language description in; parameters derived with roles and citations, so every prior has a stated source.

  2. 02

    Data

    A reference simulator generates clean calibration data — or you upload your own observations and the system scores their quality.

  3. 03

    Code

    The simulator is streamed live as it is written, grounded in 23 book-tested reference models rather than generated from scratch.

  4. 04

    Calibrate

    An identifiability screen first, then ABC-SMC: a posterior with 95% credible intervals instead of a single fitted point.

  5. 05

    Simulate

    A 30-run Monte-Carlo ensemble, in-browser, with seeds recorded per run so any trajectory can be reproduced.

  6. 06

    Validate

    Out-of-sample score, self-heal loop, and an exportable model-risk report built for review.

Zero infrastructure: the compute-heavy parts run in your browser via WebAssembly, so your data never has to leave the machine it is on.

Why AutoSim

Anyone can generate a model. We built the trust layer.

Identifiability screen

Measures which parameters your data can actually pin down — and freezes the rest, with reasons. No fake precision.

Claim classes

Synthetic data yields a plausibility claim. Real data yields a predictive claim, scored on a held-out window. The system will not let one masquerade as the other.

Honest scoring

Weighted rules that actually computed, renormalised. "Not assessed" beats a guessed number, and critical bugs cap the score.

Self-heal

An automatic repair loop that ratchets on the validity score — keeps a fix only if the score rises, reverts everything else, and stops at plateau.

Model-risk report

Methodology, posteriors, convergence, overfit ratio and limitations — regenerated byte-identically from stored metrics. Built for audit.

Eighteen months of validation engineering that an API wrapper cannot copy in a weekend.

Proof

It recovers known answers — and says so honestly.

The engine is checked against problems whose answers are known in advance, across different modelling paradigms.

β = 0.060

Planted-truth SIR: a true transmission rate of 0.06 recovered as 0.060 [0.052, 0.071] — the 95% credible interval covers the truth.

ρ = 0.70

M/M/1 queue: utilisation recovered within ±0.06 from the transient alone, and Little's law holds. Same engine, different paradigm.

72 → 84%

Self-heal on a degraded labour-market model: two kept repairs, one honest revert, an honest plateau — the whole trajectory visible to the user.

Live today

  • Full six-stage product in production
  • 500+ scenario templates across 20 domains
  • 17 chapter models from the companion book, with cited priors
  • 20 known-answer calibration benchmarks
  • Billing, quotas and audit logging already built

AlphaIQ

AutoSim is how AlphaIQ ships its method.

AlphaIQ builds calibrated, validated agent-based simulations of the markets, portfolios and institutions our clients operate in. AutoSim is the self-serve product; bespoke engagements are the hands-on path for teams who want the model built with them.

Bespoke engagementsApplicationsInsights

Applications

Illustrative examples of the questions.

Banking & market infrastructure

Tracing contagion through a counterparty network

Where an aggregate stress test reports a loss, a network simulation reports the route — and the point at which the response stops being linear.

Read→

Asset management

Crowding and unwind dynamics in a factor trade

Factor relationships hold until the population of participants changes. Simulating the participants makes regime change a mechanism rather than a break in a backtest.

Read→

Policy & institutional research

Policy transmission through heterogeneous balance sheets

The aggregate effect of a rate change is an average of very different responses. The distribution of those responses is usually the actual policy question.

Read→
All applications

Insights

Notes from the modelling bench.

All insights

Next step

Bring a system you need to defend.

Describe it in plain language and AutoSim will build, calibrate and score a simulation of it while you watch. If the question needs more than the product alone, AlphaIQ will build it with you.

Talk to AlphaIQ