Interpretable Systematic Risk around the Clock

This paper presents a novel framework that combines high-frequency market data with LLM-classified news narratives to identify interpretable around-the-clock systematic jump risks, demonstrating that macroeconomic-driven jumps command the highest risk premium and can be effectively exploited through a real-time factor-mimicking portfolio.

The Gist

Imagine the stock market as a giant, 24-hour roller coaster. Most people only watch the ride during the day when the park is open (9:30 AM to 4:00 PM). They see the big drops and the sudden loops and try to figure out why the ride jolted.

But this paper, written by Songrun He, argues that the most important jolts actually happen when the park is closed—overnight, while the lights are off and the ride is empty.

Here is the simple breakdown of what the paper does, using some everyday analogies:

1. The Problem: The "Blind Spot"

For years, financial experts have studied market crashes and jumps using high-speed cameras, but they only filmed during the day. They missed the overnight hours.

• The Analogy: Imagine trying to understand why a car crashed, but you only looked at the footage from 9 AM to 4 PM. You might miss the fact that the driver fell asleep at 2 AM or that a storm hit at 3 AM.
• The Reality: The author found that about 70% of the market's big "jumps" (sudden, sharp moves) happen overnight. If you ignore this, your understanding of risk is incomplete, like trying to solve a puzzle with half the pieces missing.

2. The Tool: The "Super-Reader" AI

To figure out why these jumps happened, the author didn't just look at numbers. He used a state-of-the-art Artificial Intelligence (an LLM called Qwen) that acts like a super-reading detective.

• The Analogy: Imagine a newsroom with thousands of reporters shouting headlines every second. A human can't read them all. But this AI is like a detective who can instantly read every headline, understand the context, and say, "Aha! The market jumped because the government announced a new tax policy," or "The market jumped because a war started in a distant country."
• The Innovation: Previous methods were like using a word counter (counting how many times the word "inflation" appeared). This AI actually reads and reasons like a human, understanding the story behind the numbers.

3. The Discovery: Not All Jumps Are Created Equal

The author used this AI to sort every market jump into different "buckets" based on the news that caused it. He found that the market pays different "prices" (risk premiums) for different types of news.

• The Analogy: Think of risk like insurance. You pay more to insure your house against a hurricane than against a broken window.
• The Finding: The market pays the highest price for jumps caused by Macroeconomic News (like unemployment numbers or GDP reports). These are the "hurricanes" of the financial world.
  • Jumps caused by a single company's earnings are like a "broken window"—the market cares, but not as much.
  • Jumps caused by geopolitical tension or policy changes are somewhere in the middle.
  • Key Insight: The "Macroeconomic" bucket was the only one that consistently made investors rich over time, outperforming the general market.

4. The Strategy: The "Smart Night Watchman"

The author built a trading strategy based on this discovery.

• The Analogy: Imagine a night watchman who knows exactly which type of storm is coming. Instead of just watching the whole sky, he focuses only on the "Macro Storm" bucket. When that specific risk appears, he buys insurance (or in this case, a specific portfolio) that pays off big when that risk hits.
• The Result: This strategy, which switches gears every year to focus on the most "priced" risk, achieved a Sharpe Ratio of 0.95.
  • What does that mean? In the financial world, a Sharpe Ratio above 1.0 is considered "excellent." The standard market average was only 0.53. This means the strategy got you nearly twice the return for the same amount of risk compared to just buying the whole market.

5. Why the "Thinking" AI Matters

The author tested two versions of the AI: one that just "spits out" answers quickly (System 1 thinking) and one that "thinks" through the problem first (System 2 thinking).

• The Analogy: It's the difference between a student guessing an answer in 5 seconds versus a student who takes 5 minutes to write out the steps, check their math, and find the right solution.
• The Result: The "Thinking" AI was vastly superior. It correctly identified the cause of the market jump 97% of the time, while the "non-thinking" version only got it right 70% of the time. The "Thinking" AI's strategy made much more money.

The Big Takeaway

This paper is a wake-up call for investors and economists:

1. Don't ignore the night: The most important market moves happen while you are sleeping.
2. Context is king: It's not enough to know that the market moved; you need to know why. AI is now good enough to read the news and tell us the "why" in real-time.
3. Macro matters most: When the big economic numbers surprise us, that is where the real money (and risk) lies.

By combining high-speed data, 24-hour monitoring, and a "thinking" AI, the author has built a map that helps investors navigate the market's hidden dangers and opportunities much better than before.

Technical Summary

1. Problem Statement

The paper addresses a critical gap in financial econometrics and asset pricing: the lack of a comprehensive, around-the-clock analysis of systematic jump risk.

• The Gap: Existing literature (e.g., Aleti and Bollerslev, 2025) primarily focuses on intraday trading hours (9:30 a.m. – 4:00 p.m.), potentially suffering from omitted variable bias by ignoring the overnight period where significant systematic risk materializes.
• The Challenge: Identifying the causes of market jumps is difficult. Traditional Natural Language Processing (NLP) methods (e.g., word counts, LDA) often lack the semantic reasoning required to link high-frequency price jumps to specific, complex news narratives. Furthermore, proprietary LLMs pose replication hurdles, and standard models may suffer from lookahead bias or hallucination.
• The Objective: To construct a framework that decomposes systematic risk into interpretable, topic-specific jump components using high-frequency data and state-of-the-art Large Language Models (LLMs) to determine if these risks are priced and how they can be exploited for investment.

2. Methodology

The author proposes a three-pronged methodological framework combining high-frequency econometrics, continuous-time modeling, and advanced AI.

A. Data Construction

• Asset Panel: A high-frequency panel of S&P 1500 constituent stocks (1997–2020) using TAQ millisecond data, cleaned and aggregated to 15-minute intervals.
• Systematic Factor: A continuous market factor constructed by combining:
  • Intraday: Value-weighted returns of NYSE/NASDAQ/AMEX stocks.
  • Overnight: S&P 500 E-mini futures returns (trading nearly 24 hours) to capture price discovery when cash markets are closed.
• News Data: Dow Jones Newswires, providing precise timestamps for news releases aligned with jump events.

B. Econometric Framework (Continuous-Time Fama-MacBeth)

The paper adapts the continuous-time Fama-MacBeth regression (Aït-Sahalia, Jacod, and Xiu, 2025) to decompose risk:

1. Jump Identification: Uses a truncated variance estimator to distinguish between continuous diffusion and discontinuous jumps in the market factor.
2. Beta Estimation:
   • Continuous Betas ($\beta^C$): Estimated via rolling windows on non-jump periods.
   • Jump Betas ($\beta^J$): Estimated using a "pure-play" approach where only observations containing systematic jumps are used. This isolates the exposure of individual assets to specific jump types, leveraging the high signal-to-noise ratio of jump events.
3. Factor-Mimicking Portfolios: Constructs portfolios that isolate specific risk sources (e.g., macro jumps, policy jumps) while neutralizing exposure to other factors. These portfolios satisfy "pure-play" properties (unit exposure to the target risk, zero exposure to others).

C. LLM-Based Narrative Retrieval & Classification

The core innovation lies in using the Qwen3-235B-A22B (an open-source, reasoning-enabled LLM) to interpret the news driving jumps:

1. Narrative Retrieval (Prompt 1): The model analyzes news headlines released within a 15-minute window of a market jump to identify the specific cause, outputting a JSON with relevant news IDs and explanations.
2. Topic Definition (Prompt 2): The model classifies the retrieved narratives into five mutually exclusive economic categories:
   • U.S. Policy Actions
   • U.S. Macro Data Surprises
   • Geopolitical & Security Events
   • Corporate Earnings & Guidance
   • International Market Spillovers
3. Classification (Prompt 3): Each jump is assigned to one of these categories.
4. Robustness Checks: The author employs ChronoBERT (a model with a pre-2000 knowledge cutoff) to rule out lookahead bias and compares the reasoning LLM against a "Non-Think" (non-reasoning) version and traditional LDA models.

3. Key Contributions

1. First Around-the-Clock Decomposition: Provides the first empirical evidence that systematic jump risk is heterogeneous and that a significant portion of priced risk occurs overnight, challenging the intraday-only focus of prior literature.
2. LLM Integration in Asset Pricing: Demonstrates that reasoning LLMs can effectively retrieve causal narratives and classify jumps into economically meaningful categories, outperforming traditional NLP (LDA) and non-reasoning models in both classification accuracy and investment performance.
3. Interpretable Risk Factors: Moves beyond "black box" factor models by linking risk premia directly to specific news types (e.g., "Macro Data Surprises"), offering a transparent mechanism for understanding why risks are priced.

4. Key Results

A. Risk Premia Heterogeneity

• Macroeconomic Dominance: The Macroeconomic Jump Risk factor commands the highest premium, with an annualized return of 3.65% and a Sharpe ratio of 0.78 (outperforming the market's 0.53).
• Other Categories: Corporate and International jump risks show positive but statistically insignificant or lower premiums once macro risk is controlled for. Policy and Geopolitical risks show mixed or negative premiums.
• Heterogeneity: A Wald test rejects the null hypothesis that all risk premia are equal, confirming that the market prices different types of jumps differently.

B. The Importance of Overnight Data

• Omitted Variable Bias: Estimating risk premia using only intraday data leads to noisy, biased, and often insignificant results. For instance, the macro jump premium becomes insignificant in intraday-only models because major macro releases often occur overnight.
• Performance Gap: A real-time strategy based on intraday-only estimates yields a Sharpe ratio of 0.28, whereas the around-the-clock strategy achieves 0.95.

C. Investment Strategy Performance

• Real-Time Strategy: A strategy that annually rebalances into the jump-topic factor-mimicking portfolio with the highest historical Sharpe ratio generates an out-of-sample Sharpe ratio of 0.95.
• Alpha Generation: This portfolio delivers significant alphas relative to the Fama-French 6-factor model, indicating it captures a distinct source of risk not explained by standard factors.
• Cost Efficiency: The strategy maintains strong performance (Sharpe ~0.85–0.93) even after accounting for high trading costs (up to 50 bps), due to low turnover (annual rebalancing).

D. Value of LLM Reasoning

• Reasoning vs. Non-Reasoning: The "Thinking" (reasoning) mode of the LLM is crucial. It successfully attributes 97.2% of jumps to news, compared to 69.7% for the non-reasoning version.
• Performance Impact: The non-reasoning model's strategy yields a Sharpe ratio of only 0.39, compared to 0.95 for the reasoning model, highlighting that the ability to "think" through complex narratives is essential for accurate risk identification.
• Lookahead Bias: Using ChronoBERT (pre-2000 cutoff) confirms that the results are not driven by the model "memorizing" future events; the performance remains robust (Sharpe 0.68).

5. Significance and Implications

• Market Architecture: As exchanges move toward 24-hour trading, this paper provides the necessary framework to understand risk dynamics outside traditional hours.
• Investment Strategy: It validates the creation of "pure-play" jump risk portfolios that can be traded in real-time, offering investors a tool to hedge or speculate on specific economic narratives (e.g., betting on macro surprises).
• Methodological Shift: The paper establishes a new standard for using open-source, reasoning LLMs in quantitative finance, moving beyond simple sentiment analysis to causal narrative retrieval and classification.
• Theoretical Insight: The findings support the Intertemporal CAPM (ICAPM), suggesting that investors are compensated for jumps that alter the investment opportunity set (specifically macroeconomic shocks), rather than just continuous volatility.

In summary, the paper demonstrates that systematic jump risk is not monolithic; it is driven by distinct, interpretable news narratives, with macroeconomic surprises being the primary priced risk factor, largely concentrated in the overnight period. The integration of reasoning LLMs is shown to be a critical enabler for accurately identifying these risks and constructing superior investment strategies.