Reasoning on Time-Series for Financial Technical Analysis

This paper introduces Verbal Technical Analysis (VTA), a novel framework that bridges time-series data and natural language reasoning to generate accurate and interpretable stock forecasts by converting price data into textual annotations and conditioning a time-series model on these verbalized insights.

The Gist

Imagine you are trying to predict the weather for next week.

Most modern "AI weather forecasters" (Large Language Models) are like brilliant meteorologists who have only read books. They can tell you, "Historically, when the barometer drops and the sky turns gray, it rains." They are great at analyzing text reports, news articles, and social media chatter about the weather.

However, they are terrible at looking at the actual raw data: the squiggly lines on a graph showing temperature changes over the last hour. They struggle to say, "The temperature dropped 2 degrees in 10 minutes, so it's going to rain right now."

On the other hand, traditional "math weather models" are like super-fast computers that crunch the numbers perfectly. They can predict the rain with high accuracy, but they are mute. They give you a number (e.g., "80% chance of rain") but cannot explain why in a way a human can understand.

This paper introduces "VTA" (Verbal Technical Analysis), a new AI that is both a brilliant meteorologist and a super-computer. It can look at the raw squiggly lines of stock prices, figure out what they mean, and then talk to you about why it thinks the stock will go up or down.

Here is how it works, broken down into simple analogies:

1. The Problem: The "Silent Math" vs. The "Chatty Bookworm"

• The Chatty Bookworm (Standard AI): Good at reading news ("The CEO quit!"), but bad at math ("The stock dropped 5%").
• The Silent Math (Traditional Models): Good at math, but can't explain its logic. If you ask, "Why did you predict this?", it just gives you a number.
• The Goal: We need an AI that can do the math and explain it in plain English, using the specific language of stock traders (like "Moving Averages" or "RSI").

2. The Solution: VTA (The "Translator" Framework)

The authors built a system with three main parts, which they call Verbal Technical Analysis (VTA).

Step A: The "Translator" (Time-Series Reasoning)

Imagine you have a secret code written in numbers (stock prices). The AI first acts as a translator. It takes the raw numbers and converts them into a "cheat sheet" of text.

• Instead of just seeing 100, 102, 98, it writes: "The price went up, then down. The average price is dropping, and the momentum is weak."
• The Magic Training: The AI tries to guess the future. If it guesses wrong, it gets a "punishment" (a math penalty). If it guesses right, it gets a "reward." But here's the trick: It only gets a reward if its written explanation matches the correct math. This forces the AI to learn that to be right, it must think correctly first.

Step B: The "Math Engine" (Time-Series Forecasting)

Once the AI has written its "thought process" (the cheat sheet), it passes that text to a second part of the system: the Math Engine.

• Think of this as a GPS. The GPS doesn't just know the road; it knows the traffic conditions (the text explanation).
• The Math Engine uses the text explanation to guide its calculation. It's like a driver saying, "The GPS says there's a traffic jam ahead, so I'll take the back road." The text helps the math engine make a smarter prediction.

Step C: The "Team Huddle" (Joint Training)

The system trains both the Translator and the Math Engine together.

• If the Math Engine makes a mistake, it tells the Translator, "Your explanation was misleading!"
• If the Translator gives a bad explanation, the Math Engine says, "I can't work with this!"
• They learn to work as a team, ensuring the final prediction is both accurate (good math) and explainable (good text).

3. Why Does This Matter? (The "Expert" Test)

The researchers didn't just test if the numbers were right; they tested if the explanations made sense to real humans.

• They showed the AI's predictions and explanations to 25 real financial experts (people who work at banks like JPMorgan and UBS).
• The Result: The experts loved VTA. They said the AI's reasoning was deep, accurate, and relevant. It sounded like a seasoned trader talking, not a robot guessing.
• In contrast, other AI models either gave great numbers with no explanation, or gave great explanations that were mathematically wrong.

4. The Real-World Payoff

The paper also tested this in a "simulated stock market." They used the AI's predictions to build investment portfolios (a basket of stocks).

• The Outcome: Portfolios built using VTA's predictions made more money and were less risky (less "bumpy") than portfolios built using other AI models.
• The Takeaway: Because the AI could explain why it was buying or selling, it made better decisions, leading to real profit.

Summary Analogy

Think of the stock market as a noisy, chaotic dance floor.

• Old AI is like a DJ who only reads the playlist (news) but can't hear the crowd.
• Traditional Models are like a sound engineer who can measure the volume perfectly but can't tell you who is dancing.
• VTA is the Head DJ who listens to the crowd's rhythm (the math), understands the vibe, and then shouts out to the crowd (the text explanation) why the music is about to change, helping everyone dance better.

This paper proves that when you teach AI to "think out loud" about numbers, it becomes smarter, more trustworthy, and actually better at making money.

Technical Summary

1. Problem Statement

The paper addresses a critical gap in financial AI: while Large Language Models (LLMs) excel at analyzing textual financial reports and sentiment, they struggle to perform Technical Analysis (TA)—the interpretation of historical price data to forecast future trends.

• The Core Challenge: The task requires a domain switch. Inputs and outputs are in the time-series domain (numerical price trajectories), but the reasoning process must occur in the natural language domain to be interpretable.
• Limitations of Existing Approaches:
  • Financial LLMs: Focus on text (reports, news) and ignore raw price data.
  • Time-Series LLMs: Often modify embedding spaces to output time-series data, sacrificing natural language reasoning capabilities and interpretability.
  • Zero-Shot Reasoning: Standard LLMs perform poorly at reasoning over raw time-series inputs without specialized training.
  • Existing Hybrid Models: Previous attempts (e.g., TimeCAP) rely on external auxiliary data or produce classification labels rather than full time-series trajectories with reasoning.

2. Methodology: Verbal Technical Analysis (VTA)

The authors propose VTA, a framework that synergizes "latent thinking" (a backbone time-series model) with "verbal reasoning" (an LLM) to produce accurate and interpretable forecasts. The framework consists of three main components:

A. Time-Series Reasoning (Verbalizing the Data)

The goal is to teach an LLM to reason over time-series data using financial technical indicators.

1. Textual Annotation: Raw time-series data ($X$) is converted into textual annotations ($X'$) containing statistical summaries (min, max, mean) and financial technical indicators (e.g., SMA, EMA, RSI, MACD, Bollinger Bands).
2. Time-GRPO (Time-Series Group Relative Policy Optimization): A novel reinforcement learning pipeline is used to optimize the reasoning trace.
   • Cold-Start: Generates initial samples.
   • Rejection Sampling & SFT: Filters for reasoning traces that lead to lower Mean Squared Error (MSE) and fine-tunes the model via Supervised Fine-Tuning (SFT).
   • Reward Objective: The model is optimized using a modified GRPO objective where the reward ($r$) is the inverse MSE ($1 / \lambda \cdot \| \hat{y} - y \|^2$). This forces the LLM to generate reasoning traces that directly correlate with accurate numerical predictions.

B. Time-Series Forecasting (Latent Thinking)

A backbone forecasting model (repurposed LLM, e.g., GPT-2) is trained to capture complex low-level patterns in the time-series data.

• Cross-Modal Alignment: The model aligns time-series tokens with language embeddings using Principal Component Analysis (PCA) and Multi-Head Cross-Attention.
• Feature Regularization: A loss function ensures that the embeddings in the temporal branch and the text branch remain consistent at every transformer layer, preventing the model from losing its language capabilities while learning time-series patterns.

C. Joint Conditional Training

To combine the two components, the forecasting model is conditioned on the reasoning attributes generated by the LLM.

• Mechanism: The model is trained to produce forecasts ($\hat{y}$) based on both the raw time-series input ($X$) and the reasoning attributes ($c$, e.g., trend direction, volatility levels) extracted from the LLM's output.
• Training Strategy: A single network is trained to handle both conditional (with reasoning) and unconditional (without reasoning) forecasts simultaneously. During inference, the final prediction is a weighted combination of the conditional and unconditional outputs, controlled by a guidance scale ($s$).

3. Key Contributions

1. VTA Framework: The first framework to explicitly combine verbal reasoning (via LLMs) and latent pattern recognition (via time-series models) for financial forecasting, ensuring both accuracy and interpretability.
2. Time-GRPO: A novel RL training method that uses inverse MSE as a reward signal to optimize LLM reasoning traces specifically for time-series forecasting tasks.
3. Conditional Forecasting: A joint training mechanism that allows the time-series model to "listen" to the verbal reasoning attributes, effectively bridging the gap between natural language logic and numerical prediction.
4. Comprehensive Evaluation: Extensive testing across US, Chinese, and European markets, including portfolio optimization analysis.

4. Experimental Results

The model was evaluated on the StockNet dataset (US) and additional datasets from the Dow Jones, FTSE China A50, and EURO STOXX 50.

• Forecasting Accuracy: VTA achieved State-of-the-Art (SOTA) performance, outperforming 12 traditional time-series models (e.g., Transformer, Autoformer, TimesNet) and 2 Time-Series LLMs (TimeLLM, CALF).
  • It achieved the lowest MSE and MAE across all datasets.
  • Compared to the second-best model, VTA showed statistically significant improvements (p < 0.05).
• Reasoning Quality: Evaluated by 25 industry experts (from firms like JPMorgan, UBS, Allianz). VTA scored significantly higher than GPT-4.1 mini and DeepSeek-R1 in:
  • Depth: Use of specific financial indicators.
  • Accuracy: Correct interpretation of technical signals.
  • Relevance: Linking indicators to the forecast.
• Portfolio Optimization: Portfolios constructed using VTA forecasts achieved the highest Sharpe Ratio (1.719) among all baselines, demonstrating superior risk-adjusted returns.
• Generalization: While VTA excels in financial domains with intrinsic interpretable signals, its reasoning capability is less impactful on non-financial time-series (e.g., Healthcare/ILI, Energy) where such signals are absent, confirming the model's domain specificity.

5. Significance and Impact

• Bridging the Gap: VTA successfully solves the "domain switch" problem, proving that LLMs can be trained to reason over numerical time-series data without losing their natural language capabilities.
• Interpretability in Finance: Unlike "black-box" time-series models, VTA provides human-readable reasoning traces that align with standard financial technical analysis practices (e.g., identifying overbought conditions via RSI or trend reversals via MACD). This is crucial for regulatory compliance and trust in financial decision-making.
• Practical Utility: The model's ability to generate high-quality portfolios with superior Sharpe ratios demonstrates that its reasoning is not just theoretically sound but practically profitable.
• Future Directions: The authors suggest incorporating more complex stock characteristics (non-stationarity) and improving the alignment between reasoning and forecasting using advanced conditioning techniques.

In summary, VTA represents a significant step forward in Explainable AI (XAI) for finance, moving beyond sentiment analysis to provide interpretable, data-driven technical analysis that rivals or exceeds the performance of specialized numerical models.