Janus-Q: End-to-End Event-Driven Trading via Hierarchical-Gated Reward Modeling

Janus-Q is an end-to-end event-driven trading framework that addresses the lack of event-centric datasets and the misalignment of language models with financial objectives by unifying a large-scale annotated news event dataset with a two-stage optimization process guided by a Hierarchical Gated Reward Model, achieving superior profitability and interpretability compared to existing strategies.

The Gist

Imagine you are trying to predict the weather to decide whether to carry an umbrella.

The Old Way (Traditional Trading):
Most trading systems are like a weather app that only looks at the temperature history of the last hour. It sees the temperature rising and guesses it will keep rising. It ignores the fact that a massive storm front (a news event) is suddenly rolling in. It's trying to predict the future based purely on the past, often getting caught in the rain because it missed the big picture.

The New Way (Janus-Q):
The paper introduces Janus-Q, a new trading system that acts more like a seasoned meteorologist who reads the news, watches the sky, and understands why the weather is changing. Instead of just looking at numbers, Janus-Q reads financial news stories and treats every story as a distinct "event" that triggers a specific reaction in the market.

Here is a breakdown of how Janus-Q works, using simple analogies:

1. The Problem: The "Noise" vs. The "Signal"

Financial markets are chaotic. Prices jump around for no reason (noise), but they also jump because of big events like a company merging, a CEO getting fired, or a new law passing (signals).

• The Challenge: Old AI models are great at reading numbers but terrible at understanding the story behind the numbers. They treat a "Company Merger" and a "Minor Typos in a Report" as just two different data points, missing the fact that one is a tsunami and the other is a ripple.
• The Gap: Also, just because an AI can write a smart-sounding article about a stock doesn't mean it knows how to actually make money trading it. It's like a chef who can write a beautiful recipe but burns the dinner.

2. The Solution: The Two-Stage "Chef's Training"

Janus-Q solves this with a two-step training process, like training a new chef.

Stage 1: The Recipe Book (Data Construction)
Before the AI can cook, it needs a massive, high-quality cookbook.

• The researchers built a dataset of 62,400 financial news stories.
• They didn't just paste the text; they labeled every story with:
  • What happened? (e.g., "Risk Warning" vs. "Dividend").
  • Who is affected? (The specific stock).
  • What was the result? (Did the stock price go up or down, and by how much?).
• Analogy: Imagine a library where every book is tagged with "This story made the stock go up 5%" or "This story made it crash 10%." This teaches the AI the real-world consequences of every type of news.

Stage 2: The Taste Test (Model Training)
Now, the AI learns to trade using this data. But instead of just memorizing the answers, it learns through a special reward system called HGRM (Hierarchical Gated Reward Model).

Think of HGRM as a strict but fair head chef who grades the AI's decisions:

• The Hard Gate (The "Stop" Sign): If the AI predicts the stock will go up, but the news clearly says it will go down, the Head Chef immediately slams the brakes. No points are awarded, no matter how good the rest of the reasoning was. This prevents the AI from making dangerous, opposite bets.
• The Soft Gate (The "Consistency" Check): If the AI correctly identifies the news as a "Merger," it gets a bonus. If it calls a "Merger" a "Dividend," it gets a small penalty. This forces the AI to understand the type of event, not just the direction.
• The Reward (The "Profit"): Finally, if the AI predicts the right direction, the right event type, and the right magnitude (how much the price moves), it gets a big reward based on the actual profit it would have made.

3. Why It Works Better

Most trading AIs are like a hamster running on a wheel—they are fast but going in circles, reacting to the last second of price movement.

Janus-Q is like a chess player.

• It pauses to read the "board" (the news).
• It identifies the "piece" that moved (the event type).
• It calculates the likely outcome based on history (the CAR data).
• It makes a move that is consistent with the rules of the game (financial logic).

The Results

When they tested Janus-Q against:

1. Market Indices (Just buying the whole market).
2. Old AI Models (Time-series predictors).
3. Smart Chatbots (General AI models).

Janus-Q won.

• It made 102% more profit per unit of risk (Sharpe Ratio) than the second-best strategy.
• It was 17.5% more accurate at guessing which way the stock would move.
• Crucially, it didn't just get lucky; it made decisions that humans could understand and trust because it explained why it made the trade based on the news story.

In a Nutshell

Janus-Q is a trading system that stopped trying to predict the future by staring at a spreadsheet of numbers. Instead, it learned to read the news, understand the story behind the stock, and make trades based on the logic of the event. It's the difference between guessing the weather by looking at a thermometer and actually reading the forecast.

Technical Summary

1. Problem Statement

Financial markets are frequently driven by discrete, interpretable events (e.g., earnings announcements, M&A, regulatory changes) rather than smooth temporal dynamics. Existing learning-based trading systems face two critical limitations:

• Data Limitations (Ch1): There is a lack of large-scale, event-centric datasets that jointly model news semantics (event type, sentiment) with statistically grounded market reactions (Cumulative Abnormal Returns, CAR). Most existing datasets are either too coarse, domain-specific, or fail to link specific events to their quantitative economic impact.
• Model Misalignment (Ch2): Current approaches often treat textual news as an auxiliary input to numerical time-series models or rely on Large Language Models (LLMs) that generate plausible reasoning but lack alignment with realized market outcomes. Purely supervised learning captures superficial correlations, while standard Reinforcement Learning (RL) often relies on heuristic, linearly additive rewards that fail to capture complex economic trade-offs (e.g., risk vs. return, event-type consistency).

2. Methodology: The Janus-Q Framework

Janus-Q proposes a two-stage, end-to-end event-driven trading framework that elevates financial news events from auxiliary signals to primary decision units.

Stage I: Event-Centric Data Construction

The authors constructed a large-scale dataset comprising 62,400 expert-annotated financial news articles covering 5,282 Chinese A-share stocks.

• Annotation: Each event is labeled with 10 fine-grained event types, associated stocks, semantic polarity, and trading strength.
• Market Impact Quantification: The dataset utilizes Event-to-CAR modeling. Using a standard event study methodology, the authors calculate Cumulative Abnormal Returns (CAR) over a defined event window.
  • Market Model (MR): Removes broad market movements using a benchmark index.
  • Risk Model (RM): Further neutralizes systematic style and industry effects using a multi-factor risk model (Barra-style CNE5) to isolate idiosyncratic, event-driven returns.
• Labeling: The ground truth label $y = \{c, d, s, e\}$ includes the realized CAR ($c$), direction ($d$), trading strength ($s$), and event type ($e$).

Stage II: Decision-Oriented Fine-Tuning

The model undergoes a multi-step optimization paradigm to align semantic reasoning with executable trading actions.

1. Supervised Fine-Tuning (SFT): Initializes the model to learn a reasoning-aware mapping from event descriptions to expected CAR, integrating textual semantics with market signals.
2. Reinforcement Fine-Tuning (RFT) with HGRM: The core innovation is the Hierarchical Gated Reward Model (HGRM), which guides Group Relative Policy Optimization (GRPO). Unlike standard RL, HGRM decomposes the reward into interpretable components with a hierarchical structure:
   • Hard Gate (Direction): If the predicted direction contradicts the realized market movement, the reward is blocked ($g_{dir}=0$), preventing spurious profits from incorrect polarity.
   • Soft Gate (Event-Type Consistency): A multiplicative discount factor ($m_{evt}$) reduces rewards if the predicted event type does not match the ground truth, ensuring the model understands why a trade is made.
   • Trading Reward (PnL): A cost-aware payoff based on realized CAR, scaled by trading strength and the event-type discount.
   • Magnitude Shaping & Process Reward: Encourages accurate CAR magnitude prediction and penalizes excessive self-questioning or missing reasoning steps.

3. Key Contributions

1. Novel Dataset: Creation of a unified benchmark for event-level market impact analysis, linking 10 event types with semantic labels and precise CAR calculations for 62,400 events.
2. Janus-Q Framework: The first end-to-end event-driven trading framework that directly maps financial news events to trading decisions, bypassing the intermediate step of numerical time-series forecasting.
3. Hierarchical-Gated Reward Modeling (HGRM): A novel reward mechanism that explicitly models trade-offs between direction accuracy, event-type consistency, and profit magnitude, ensuring the policy learns financially grounded reasoning rather than exploiting noise.

4. Experimental Results

The framework was evaluated on a test set from November 2024 to February 2025, comparing against market indices, time-series LLMs, financial domain LLMs, and general-purpose LLMs.

• Trading Performance: Janus-Q achieved a Sharpe Ratio of 1.3088, outperforming the runner-up (QwQ-32B) by 102.0%. While most baselines and market indices showed negative Sharpe Ratios during the test period (characterized by market correction), Janus-Q maintained a stable, consistent growth trend.
• Decision Accuracy:
  • Direction Accuracy (DA): Improved by 17.5% over the strongest competing strategies.
  • Event Type Accuracy (ETA): Improved by 18.2%.
  • Prediction Error: Reduced Mean Absolute Error (MAE) by 18.3% compared to the best time-aware model and 50.6% compared to the best financial LLM.
• Ablation Studies:
  • Removing Supervised Fine-Tuning (SFT) caused the Sharpe Ratio to drop to -5.28, highlighting its role in establishing a stable reasoning foundation.
  • Removing the HGRM components (specifically direction and magnitude rewards) significantly degraded performance, confirming the necessity of the hierarchical reward structure.
• Human Alignment: In a case study with human experts (analysts and CFA holders), Janus-Q showed high agreement in event interpretation, with tie rates of 40.5%–52.0% and loss rates below 25.0% against human judges.

5. Significance

Janus-Q demonstrates that explicitly modeling financial events as primary decision units is superior to traditional time-series approaches. By bridging the gap between unstructured text and quantitative market outcomes through a structured, hierarchical reward system, the framework achieves:

• Interpretability: Decisions are driven by explicit event reasoning rather than black-box pattern matching.
• Robustness: The model adapts effectively to shifting market regimes where purely numerical models fail.
• Profitability: It translates semantic understanding into tangible, risk-adjusted trading gains, proving that aligning LLM reasoning with financial objectives is a viable path for next-generation algorithmic trading.