AI in finance has gone from buzzword to boardroom necessity. But let’s be honest—most articles on the topic either sound like they’re written by robots or aimed at them. So, here’s a human take on the AI revolution shaking up quantitative finance. No jargon bingo, no empty hype—just practical insight with a bit of cheek.
1. The Two AI Camps: Logic vs. Language
On one side, we have symbolic engines—Wolfram Alpha, SageMath, the maths world’s equivalent of a pocket protector. Brilliant with equations, less so with anything resembling natural conversation.
On the other, LLMs like GPT-4o and Gemini. These models read like humans, talk like humans, sometimes even argue like humans—but don’t always count like them. They’re more Shakespeare than spreadsheet when it comes to precision.
The trick? Use both. LLMs for thinking, engines for calculating. It’s like having a mate who writes the pub quiz questions and another who checks your answers.
2. The Math Olympics: Where AI Shines—and Faceplants
LLMs are smashing high-school level benchmarks. They’re the top of the class at understanding logic, structure, and how to spin a good tale around numbers.
But throw them into the research-level ring—think FrontierMath—and they fall flat. Less genius, more GCSE. The takeaway? AI isn’t here to invent new maths, it’s here to make applying what we know faster, cleaner, and cheaper.
3. AI Tools: Meet Your Quant Team’s New Best Mates
| Tool | Best At | Not Great At |
|---|---|---|
| GPT-4o | Breaking down word problems, writing code | Numbers that don’t lie |
| Gemini | Handling huge documents, Google Sheet magic | Slight LaTeX hiccups |
| Wolfram Alpha | Bulletproof calculations | Needs clear, formal input |
| SageMath | Complex models | Steep learning curve |
4. Alpha, But Smarter: How AI Powers Your Investment Strategy
From scanning tweets for sentiment to spotting patterns in old market data, AI is redefining the quant workflow:
- Pre-Trade: NLP for news, ML for pattern-hunting.
- Execution: Algorithmic trading with brains and brawn.
- Post-Trade: Real-time risk modelling with a side of anomaly detection.
Add deep learning for prediction, reinforcement learning for portfolio tweaks, and LLMs that can summarise analyst reports in plain English—and you’ve got yourself an alpha pipeline worth bragging about.
5. Prompt Engineering: The Secret Sauce
If AI is the engine, the prompt is the steering wheel. And bad prompts? Like texting while driving.
Want your LLM to spit out a Discounted Cash Flow model? Don’t say “build me a model.” Say:
“Act as a financial analyst. Calculate NPV for a project with a £200k outlay and cash inflows of £50k, £60k, £70k, at 10% discount rate.”
Even better—think of it as a conversation. Break it down. Tell it what role to play. Specify outputs. Think like a CFO talking to a very clever intern.
6. Risks: When AI Hallucinates Profits
Here’s where things get dicey. LLMs sometimes make stuff up. Confidently. And with charts.
From fictional stock prices to quoting court cases that never happened, AI hallucinations are real. In finance, they’re also expensive.
The fix? Use Retrieval-Augmented Generation (RAG), domain-specific training, and human-in-the-loop reviews.
- RAG means grounding the AI in real, trusted data—think feeding it company reports instead of letting it guess from internet noise.
- Domain-specific training involves fine-tuning the AI on specialist financial content, so it speaks your language (literally and numerically).
- Human-in-the-loop reviews ensure a qualified pro checks the AI’s homework before any decisions are made. Trust, but verify. Always.
7. Looking Ahead: The Rise of the AI Copilot
The future of finance isn’t AI running the show—it’s AI riding shotgun. The pros use it to:
- Write code faster
- Turn messy data into insights
- Summarise complex reports
Eventually, we’ll have autonomous agents drafting pitchbooks, spotting anomalies, even drafting strategy. But for now? It’s about doing what you already do, just much, much faster.
My Final Tip
If you want to test an LLM’s maths skills, ask it to solve something like 8.8 – 8.11. Sounds easy, right? Not for AI—most models fumble it due to floating-point precision issues. The trick? Prompt it to use rational arithmetic in Python and it’ll nail it.
We also tested this classic finance prompt—“Work out compound interest on a £500k pension at 5% over 10 years”—across ChatGPT o3 Pro, Claude Sonnet 4, Gemini 2.5 Pro, and Meta AI. All delivered a clear, step-by-step method and agreed on the final answer. So yes, with the right question, AI can show its working beautifully.
