LLMs can predict the next word. They can't predict the next consequence.
Ask a language model what a closed Strait of Hormuz does to oil. It answers instantly and correctly. It has read that sentence a thousand times. Ask what it does to a sugar producer in Brazil. Same fluency. Now it is guessing.
Language models learn which words follow which. They don't know why things happen or what plausibly follows. When they don't know, they invent causes and effects, convincingly. More scale does not fix this.
On CausalProbe (2024), a frontier model scored below 70 % on hard causal questions. On EconCausal (2025), accuracy fell 32.6 % when the economic context shifted. In trading, a missed effect is a position on a phantom signal.
In finance, trust is the product.
When an AI can't explain its reasoning, you can't trust its conclusions. An AI that mistakes correlation for cause gets it wrong, convincingly. In markets, that is expensive.
Today's finance AI traces a number back to the filing it came from. It never checks the economic logic that number sets in motion. A citation verifies an input, not an inference.
Regulation pushes the same way. The EU AI Act, on top of DORA and MiFID, raises the bar for explainable, auditable AI in financial workflows. When the investment committee asks why the model suggested a position, "it was in the training data" won't do. A chain of sourced, dated evidence will.
| Sentiment analysis | Too shallow. A sign, not a mechanism. |
| Correlation models | Blind to a changing world. They break when the regime changes. |
| Generic LLM workflows | Fluent but inconsistent. Weak past the first effect. They cite inputs, not inferences. |
| Human analysts | Slow. They catch the first effect and miss the second and third. |
One event. Two assets. Opposite signs.
28 February 2026: the U.S. and Israel strike Iran. What does that do to two ethanol producers, one in Brazil, one in the United States? A worked example of the reasoning, not a realised prediction.
Same cause, two paths, opposite signs. LLMs cannot trace back both paths to the same cause and realise their opposing effect signs.
Change one thing in the world and the U.S. leg flips. Everything upstream stays the same. LLMs cannot causally predict what-if scenarios, hence the consequences of interventions.
We don't forecast markets.
We predict what an event sets in motion: which assets, which way, and why.
AI is moving into financial decisions, and the systems that understand cause and effect in the economy will be the infrastructure underneath. Ours is a live causal model of that economy: actors, assets and countries, linked by causal relationships checked against evidence. Give it an event and it traces the chain to the assets that will move.
We read the world ourselves. Our own near-real-time observer of the financial press keeps the model current as news breaks.
Event → asset impact
Which assets move, in which direction, through which chain, and with what confidence.
Scenario modelling
The most likely outcomes of an event, and how they cascade through sectors and markets.
What-if and counterfactual
"If X happens, then…" Test a price cap, a sanction, an outage before it happens. Or ask what would have happened without it.
- Causal chain
- The steps of cause and effect from an event to a price. Strikes, a closed strait, dearer gas, dearer fertilizer.
- Multi-hop
- Following that chain several steps down, not stopping at the first. Sugar is four steps from the strikes.
- Intervention
- Someone changes something in the world. A regulator caps a price. Everything downstream changes with it.
- Counterfactual
- What would have happened without that change. It separates what an event caused from what was already moving.
FINCH
FINancial Claims Hypergraph
FINCH reads financial news and turns it into a live graph of causal claims: who affects whom, with every link carrying its source, its date and its market context. Ask it a question and it follows the claims instead of guessing, then cites the articles it stood on.
Explore the graph in Ampiric Universe, our 3D view of it. Access is by invitation while FINCH is in evaluation.
A complement to the LLM stack, not a rival.
The finance copilots and terminals people already use read, summarise and draft well. They don't reason about consequences. We ground them: the language model reads and writes, our causal model traces what an event sets in motion, with evidence on every link.
- Traces the chain several steps down: oil, LNG, fertilizer, corn, ethanol, and where they meet.
- Attaches a sourced, dated citation to every link, not just to the inputs.
Checked chains, not confident guesses.
Causal AI for finance is not a new idea. What we own grows every day: a causal model of the world economy, fed by our own reader of the financial press, with dated evidence on every link. And a team that has built data pipelines at scale before.
Checked inference, not just cited input
Others trace a number to its source. We check the causal link between events against evidence, and show the chain.
Same cause, opposite effects
One event, two assets, two signs, and the reason. Several steps down the chain, across sectors. Pattern matching can't do that.
What-ifs and counterfactuals
A price cap, a sanction, an outage: what changes? What would have happened without the event? These are the questions language models fail on.
Traceable by construction
Every link is backed by sourced, dated evidence. The chain is the audit trail an investment or risk committee can check.
Built for the desks where a wrong "why" is expensive.
For hedge funds, asset managers, commodity desks and bank research. Three applications. Each plugs into a workflow you already run.
Event-driven trading
An event breaks. Minutes later you know which assets move, which way, and why.
Event impact
Event in. Affected assets out, with direction, chain and confidence.
Second-order effects
A shock moves more than oil. Shipping, fertilizer, currencies, some the other way. Everyone else stops at the first effect.
Signals for quants
Impact features for your own models, over an API. No prose.
Scenario analysis
"If X happens, what happens to my book?" One traceable cascade instead of a whiteboard.
What if they cap the price?
Change one thing in the world. We re-trace every path and show which outcomes flip.
What if it hadn't happened?
We separate what the event caused from what was already moving. Your thesis is judged on the right evidence.
Hormuz closed for three weeks?
A named shock run against your positions. Every exposure comes with its reason.
Risk & audit
An event changes your exposure. You see it, with the reason, before the committee asks.
Exposure flags
A new event, the holdings it touches, the direction, the reason.
The audit trail
Every conclusion carries its chain and its sources. Something a committee can check.
Explainability on record
The EU AI Act, DORA and MiFID raise the bar. The reasoning behind each conclusion is on file.
Delivered as an API and a copilot, inside the tools you already use.
Finance is the proving ground, not the limit.
Finance first, because it is fast, noisy, rich in data and quick to tell you whether you were right. The same engine, with a different vocabulary, serves every domain where events ripple through chains nobody sees whole.
Finance
Event-driven trading and cross-asset market intelligence.
- Hedge funds and asset managers
- Commodity and energy trading desks
- Bank research and strategy teams
Enterprise
Commodity and supply-chain shocks. Disruption monitors, input-cost and price forecasting for chains nobody models live today.
- Energy-intensive companies
- Supply-chain and procurement teams
- Risk and treasury teams
Strategy
Geopolitical and policy scenarios. Sanctions and board scenarios are manual and subjective today. Tools that trace the consequences.
- Government agencies
- Defence
- Global enterprises
Ampiric: amplified empiricism.
The name is a commitment. We test our ideas against reality. We measure what matters. We build technology that works in the real world, not just in benchmarks.
Ampiric began as a massive, custom effort to read the world's financial press in near real time. Then we saw the bigger idea. The value is not in reading more news. It is in reasoning about what the news sets in motion.
AI expertise. Production discipline.
Four PhDs and a production ML engineer. AI, control systems, industrial data pipelines, knowledge representation. Ten to fifteen-plus years each, in industry.
PhD in AI. 10+ years of research and product development, previously director-level, managing large engineering teams.
PhD in control systems. 15+ years developing data analytics in high-tech industry, turning complex, noisy signals into decisions.
PhD in industrial engineering. 15+ years of software development, as tech lead and architect for large-scale data pipelines.
PhD in knowledge representation. Large-scale recommender systems, algorithmic fairness and structured knowledge retrieval.
Software engineer. ML research in speech, sentiment recognition and computer vision; scales ML pipelines in production.
We're building — talk to us.
We work with a small circle of design partners and pilot users, and we are talking to investors. If you run a desk where a wrong "why" is expensive, or you back teams that build infrastructure, write to us.
Or write to us directly: [email protected]