From Deep Learning to LLMs: How AI Is Rewriting the Quant Playbook
https://arxiv.org/abs/2503.21422
If you think quantitative investing is all about math geeks staring at spreadsheets, think again. Over the past decade, AI has been quietly and now loudly reshaping how money is managed from how we predict markets to how trades are executed.
A new survey paper, From Deep Learning to LLMs: A Survey of AI in Quantitative Investment maps out this transformation in detail.
Why Quant Investing Needs AI
Quantitative investment (“quant”) means using algorithms, data, and systematic rules to make investment decisions. Instead of relying on human hunches, quants look for patterns in vast oceans of data to find “alpha” returns above the market average.
Traditionally, this involved:
Manually crafted features — human analysts decided which variables might predict stock movements.
Statistical models — think regressions and factor models.
Rule-based execution — fixed trading rules to place orders.
But as markets grew faster, noisier, and more interconnected, these methods started to hit a ceiling.
Stage 1 → Stage 2: The Deep Learning Revolution
Deep Learning (DL) changed the game by:
Seeing patterns humans can’t — from subtle correlations in price movements to hidden sentiment shifts in news.
Scaling across the whole pipeline — from cleaning raw data, to predicting prices, to deciding how much to buy or sell.
Handling messy inputs — like text, images, graphs, and alternative data (social media posts, satellite images, even WiFi signals).
Example: DL models can blend candlestick charts, supply chain links, and Twitter sentiment into a single prediction signal.
But… DL models have two big problems:
They can overfit (look smart in backtests but fail in real life).
They’re hard to explain (the “black box” problem).
Stage 2 → Stage 3: Enter the LLM Era
Large Language Models (LLMs) like GPT-4 are taking AI in quant from AI-powered to AI-automated.
How?
Processing unstructured data: earnings call transcripts, analyst reports, Reddit threads.
Generating new trading ideas: turning qualitative insights into quantitative factors.
Acting as autonomous agents: chaining together prediction, portfolio optimization, and order execution.
For example, an LLM-powered “quant agent” could:
Read financial news in real time.
Detect a shift in sentiment around a stock.
Recalculate portfolio weights.
Place trades. All without human intervention.
The Alpha Pipeline: A Four-Step Engine
Whether powered by humans, DL, or LLMs, the goal is the same: maximize returns while managing risk.
Data Processing — Clean, unify, and engineer features from market data.
Prediction — Forecast future returns, volatility, or price moves.
Portfolio Optimization — Decide how to allocate capital across assets.
Order Execution — Place trades efficiently to minimize cost and market impact.
LLMs are starting to plug into every step, not just prediction.
The Road Ahead: Opportunities & Challenges
What’s exciting:
AutoML for finance — automating model tuning for faster adaptation.
Explainable AI — making predictions transparent for risk managers.
Knowledge-driven AI — blending human expertise with machine learning.
End-to-end agents — going straight from raw data to executed trades.
What’s tricky:
Aligning LLM “sentiment” with real market reactions.
Handling low-latency, high-frequency decision-making.
Integrating with institutional-grade risk and execution systems.
Why This Matters
We may be entering an era where trading desks run on teams of human-AI hybrids, with LLM agents doing the heavy lifting and humans focusing on strategy, oversight, and creative alpha ideas.
As the paper’s authors put it, the shift from deep learning to LLMs isn’t just an upgrade. It’s a paradigm change. The race is on to see who can best combine raw computational intelligence with market-savvy execution.
If you’re a trader, data scientist, or just curious about where AI meets money, keep an eye on this space. The first firms to master this AI-quant fusion may set the investment playbook for the next decade.