💰 quantitative finance

Quantformer: from attention to profit with a quantitative transformer trading strategy

This paper introduces Quantformer, a transformer-based neural network enhanced by transfer learning from sentiment analysis, which effectively captures long-range dependencies and market signals to outperform traditional 100-factor strategies in predicting stock returns within the Chinese capital market.

Original authors: Zhaofeng Zhang, Banghao Chen, Shengxin Zhu, Nicolas Langrené

Published 2026-02-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Zhaofeng Zhang, Banghao Chen, Shengxin Zhu, Nicolas Langrené

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict the weather to decide whether to carry an umbrella. Traditional methods might look at yesterday's rain and guess today will be wet. But the stock market is like a chaotic, stormy ocean where the weather changes every second, influenced by millions of tiny ripples, giant waves, and even the mood of the sailors.

This paper introduces a new tool called Quantformer. Think of it as a super-smart, weather-forecasting AI that doesn't just look at the rain; it understands the entire storm system to predict exactly where the sun will shine next.

Here is the story of how they built it, explained simply:

1. The Problem: The Old Tools Were Too Simple

For decades, investors have used "factors" to pick stocks. Think of a factor as a single clue, like "Is the company making money?" or "Is the stock price going up?"

The Old Way: Imagine trying to navigate a maze by only looking at one wall at a time. You might get lucky, but you'll likely hit a dead end. Traditional math models (like the ones used in the 1950s) are too rigid. They can't handle the messy, emotional, and fast-changing nature of the stock market.
The Machine Learning Attempt: Recently, people tried using AI (Machine Learning) to fix this. They used models designed for reading text (like translating English to French) to read stock charts. But this was like trying to read a book using a dictionary meant for a different language. The numbers didn't fit the "words" the AI expected.

2. The Solution: Enter "Quantformer"

The authors took a famous AI architecture called the Transformer (the same brain behind tools like ChatGPT) and gave it a makeover specifically for money.

The Metaphor: Imagine a translator who usually translates books (text). The authors realized that stock data isn't a book; it's a symphony of numbers.
The Makeover:
- No More "Word Embeddings": In normal AI, words are turned into "word vectors" (like assigning a color to every word). Since stock data is just numbers (prices, volume), the authors replaced the "word dictionary" with a simple linear layer. It's like swapping a complex thesaurus for a straight ruler. It measures the numbers directly.
- Removing the "Mask": In language AI, the model often has to guess the next word one by one (like filling in a crossword). In trading, you don't need to guess word-by-word; you need to see the whole picture of the next month's return at once. So, they removed the "mask" that forced the AI to guess step-by-step.
- The "Attention" Superpower: The core of the Transformer is "Attention." Imagine you are at a noisy party. A normal person hears everything equally. A Transformer with "Attention" can instantly tune out the chatter and focus only on the one person shouting the most important news. Quantformer uses this to look at 20 days of past stock data and instantly figure out which days actually matter for the future.

3. How They Tested It: The "Giant Simulation"

They didn't just guess; they ran a massive simulation.

The Data: They fed the AI data from 4,601 Chinese stocks over 13 years (2010–2023). That's over 5 million data points. It's like training a chef by having them taste every dish in a massive buffet for a decade.
The Training: They taught the AI to look at the past (prices and how many shares were traded) and predict the future (will the stock go up or down?).
The Competition: They pitted their new AI against 100 other traditional trading strategies. It was like a race between a Formula 1 car (Quantformer) and 100 bicycles (traditional factors).

4. The Results: The AI Wins the Race

The results were impressive:

Higher Profits: The Quantformer strategy made significantly more money than the 100 traditional strategies. While the traditional ones often lost money or broke even, Quantformer delivered strong annual returns (around 17% in their best monthly test).
Better Safety: In investing, it's not just about how much you win, but how much you risk losing. Quantformer had a lower "Value at Risk" (VaR). Think of this as the size of the hole you might fall into. Quantformer's hole was smaller and safer than the others.
The "Alpha": In finance, "Alpha" is the score for being smarter than the market. Quantformer had a positive Alpha, meaning it genuinely outsmarted the market, not just got lucky.

5. Why This Matters

The paper shows that we don't need to treat stock markets like text to use powerful AI. By tweaking the AI to understand numbers instead of words, we can build a system that:

Sees the big picture: It connects long-term trends with short-term spikes.
Adapts: It learns from the chaos of the market better than old math formulas.
Makes Money: It turns complex data into profitable decisions.

In a nutshell: The authors took a super-intelligent language brain, taught it to speak "Finance," and showed that it can navigate the stock market's stormy seas much better than the old compasses we've been using for decades. It's a bridge from "Attention" (the AI's focus) to "Profit" (the investor's goal).

1. Problem Statement

Traditional quantitative trading faces significant challenges in navigating complex, dynamic financial markets. Existing methods struggle to:

Capture Long-Range Dependencies: Traditional models often fail to effectively model long-term market trends and complex interdependencies between variables.
Handle Numerical Inputs: While Transformer architectures (originally designed for NLP) excel at capturing long-range dependencies, they rely on word embeddings and positional encodings designed for categorical text data. Financial data, however, consists primarily of numerical time series (e.g., returns, turnover rates) and categorical metadata (e.g., industry sectors), which cannot be directly processed by standard NLP embeddings.
Adapt to Sequence-to-Point Tasks: Standard Transformers are often built for sequence-to-sequence (seq2seq) tasks (e.g., translation). Stock prediction, however, is typically a sequence-to-point task where the goal is to forecast a single future value (return) or probability distribution based on a historical sequence, making the standard decoder and masking mechanisms inefficient or unnecessary.

2. Methodology: The Quantformer Architecture

The authors propose Quantformer, a modified Transformer architecture specifically adapted for quantitative financial data.

A. Data Preprocessing

Input Features: The model uses a rolling window of historical data (20 time steps). For each stock, the input matrix $X \in \mathbb{R}^{20 \times 2}$ $X \in R^{20 \times 2}$ consists of:
1. Accumulated Daily Return: $r = (P_t - P_{t-1}) / P_{t-1}$ .
2. Accumulated Daily Turnover Rate: Reflecting trading volume and investor sentiment.
Normalization: Inputs are normalized using Z-score (zero-mean, unit-variance) to handle outliers and ensure feature comparability.
Target Labels: Instead of predicting raw prices, the model predicts the empirical quantile of future returns. The target is a one-hot vector representing whether a stock falls into the top, middle, or bottom quantiles (e.g., top 20%, middle 20%, bottom 20%) of the return distribution.

B. Architectural Modifications

Quantformer modifies the standard Transformer (Vaswani et al., 2017) in three key ways:

Linear Embedding (Replacing Word Embedding):
- Standard Transformers use an embedding layer to map discrete tokens to vectors.
- Quantformer replaces this with a standard linear layer ( $X' = XW + \theta$ ). This allows the model to directly process continuous numerical data without the need for discrete tokenization.
Removal of Positional Encoding:
- In NLP, position is arbitrary and must be explicitly encoded.
- Quantformer argues that in time series data, the temporal order is inherently embedded in the sequence structure. Therefore, the positional encoding module is removed to reduce noise and overfitting.
Simplified Decoder (Sequence-to-Point):
- The autoregressive decoder and masking operations (used for generating token sequences) are removed.
- The model outputs a single probability distribution over the future return quantiles directly from the encoder output, treating the task as a classification problem.

C. Training and Strategy

Loss Function: Mean Squared Error (MSE) is used to minimize the distance between predicted and actual probability distributions.
Trading Strategy:
- Stocks are ranked based on the predicted probability of being in the top quantile.
- A portfolio is constructed by selecting the top $b$ quantile groups (e.g., top 20%).
- The strategy is tested on the Chinese A-share market (Shanghai and Shenzhen exchanges) from 2010 to 2023.

3. Key Contributions

Novel Architecture for Finance: The introduction of Quantformer, the first Transformer variant specifically designed to handle raw numerical time series without word embeddings or positional encodings.
Transfer Learning from Sentiment Analysis: The authors leverage the Transformer's ability to capture complex dependencies (originally developed for NLP sentiment analysis) and successfully transfer this capability to quantitative factor construction.
Comprehensive Empirical Validation: The study utilizes a massive dataset of 4,601 stocks over 14 years (2010–2023), covering more than 5 million rolling data points.
Benchmarking: The model is rigorously compared against 100 traditional price-volume factors and various trading frequencies (daily, weekly, monthly).

4. Experimental Results

The study evaluated the model across different frequencies and training scales.

Performance vs. Traditional Factors:
- The Quantformer factor (Monthly frequency, selecting top 20%) achieved an Annual Return (AR) of 17.35% and an Annual Excess Return (AER) of 19.43%.
- Sharpe Ratio: 0.915, significantly outperforming the average of 100 traditional factors (which had a negative average Sharpe ratio of -0.36).
- Alpha: Positive alpha (0.162), indicating the strategy generates returns above the market benchmark (CSI 300) after risk adjustment.
- Risk Management: The 99% Value at Risk (VaR) was 2.81%, lower than most traditional factors, indicating better downside protection.
Frequency Analysis:
- Monthly: Best overall performance (Highest AR and Sharpe).
- Weekly/Daily: Showed mixed results, with daily strategies having lower win rates (~44%) but consistent turnover.
Training Scale Sensitivity:
- The model remained robust when trained on different quantile definitions (selecting top 1%, 5%, 10%, or 20% of stocks).
- The QF 1% strategy (selecting only the top 1% of stocks) achieved the highest AR of 24.71% and AER of 35.74%, though with higher volatility.

5. Significance and Implications

Bridging NLP and Finance: The paper demonstrates that Transformer architectures can be effectively adapted for non-textual, numerical financial data, opening new avenues for applying large-scale deep learning models in quantitative finance.
Superior Factor Construction: By integrating market sentiment (via turnover rates) and price dynamics into a unified attention mechanism, Quantformer creates a more robust investment factor than traditional linear or static models.
Practical Applicability: The strategy is designed for real-world constraints (e.g., no short-selling in China, transaction fees) and shows consistent profitability over a long historical period, suggesting it is a viable tool for portfolio optimization.
Open Source: The authors provide the implementation code on GitHub, fostering reproducibility and further research in ML-based quantitative trading.

In conclusion, Quantformer represents a significant step forward in quantitative trading by successfully adapting the state-of-the-art Transformer architecture to the specific constraints and data types of financial markets, outperforming both traditional factor models and other machine learning approaches in terms of risk-adjusted returns.