Following the August 1999 Federal Open Market Committee (FOMC) announcement, the S&P 500 Index did not instantaneously settle at a new value. Instead, as shown in Figure 1, its price swung in both positive and negative directions for over an hour.

This example highlights a key question: how much time do financial markets need to fully react to monetary policy announcements? Applied macroeconomists often measure these price changes within narrow windows around news events such as monetary policy announcements. Small windows avoid contamination from unrelated news, but assume that the market's first response is its final response. On the other hand, wide windows allow more time for the market to react, but risk mixing the impact of monetary policy announcements with other information. Despite these trade-offs, the monetary policy literature has often defaulted to an “ad-hoc” 30-minute window. This “one-size-fits-all” approach is problematic because it assumes all markets take the same amount of time to react to monetary policy news, even though markets for longer-maturity assets might need more time to understand its more complex information. If the chosen window is wrong, measurements of monetary policy surprises and shocks can become attenuated and less relevant.

I propose a systematic method that pins down these “optimal” event windows. First, I train a neural network for text analysis to approximate the underlying relationship between the text of monetary policy announcements and the market price reaction. This process creates a “text-based signal”, which represents the network's predictions of the market price response based only on the FOMC statements, word-for-word. Second, I have the network learn this relationship and generate this text-based signal for different event window lengths (e.g., a 10-minute window, 20-minute window, and so on). The “optimal” window is defined as the length where the neural network can make the most accurate predictions. This “Goldilocks” window is the time where the market has fully reacted to the monetary policy announcement, but not so much time that the price becomes overwhelmed by noise.

My method yields four key findings. First, regardless of maturity, markets require at least 40 minutes to fully react, an event window beginning 10 minutes before and at least 30 minutes after statement release. Second, Figure 2 below shows that this optimal window length increases with underlying asset maturity, reaching 50–60 minutes for long-horizon contracts.

Third, timing differences significantly alter forward guidance surprises, with correlations for fifteen-year-ahead assets declining by 10%. Finally, adopting optimal windows corrects for the attenuated impact of shocks on yields and equity prices, resulting in macroeconomic responses that are nearly 10% more precise.

The State of Event Window Lengths

This paper joins an investigation into appropriate event window length. While early work by Hillmer and Yu (1979) suggested adjustments could last hours, the high-frequency era settled on the 30-minute window (Gürkaynak, Sack, and Swanson, 2005) Recent work, such as Casini and McCloskey (2025), provides a framework for identifying when a policy surprise “dominates” market noise, suggesting the optimal event window is an empirical parameter rather than a constant. I advance the field of natural language processing in economics by using the context-aware XLNet-Base neural network to capture semantic nuances in Fed communication (Handlan, 2022; Piller, Schranz, and Schwaller, 2025), using the text itself to solve the key timing problem.

Systematically Estimating Optimal Event Windows

I analyse 165 scheduled FOMC statements from May 1999 to October 2019. The model's inputs are the word-for-word texts, pre-processed to focus strictly on the economic outlook. The outputs consist of price log-differences constructed from high-frequency tick data for Federal Funds, Eurodollar, and Treasury futures (up to 30-year), alongside the S\&P 500 Index and its E-mini futures. Price log-differences are constructed for event windows in 10-minute intervals, with the beginning price always 10 minutes before the release of an FOMC statement.

To determine the optimal window, I approximate the non-parametric relationship between statement text and price discovery using XLNet-Base, a transformer-based neural network. Unlike other neural networks, XLNet-Base uses a permutation-based pre-training phase, allowing it to learn context bi-directionally without using the common “masking” technique, which assumes word independence and can create training discrepancies. This feature makes XLNet-Base ideal for nuanced central bank language. I train the network to predict price changes based only on text across various window lengths up to 70 minutes.

The “optimal” window is the length that maximises out-of-sample predictive performance averaged across the sample splits (this is measured as the averaged out-of-sample coefficient of determination). This represents the “Goldilocks'” window: the duration where the text-based signal is most precise, indicating the market has processed the information while the relative impact of noise is minimised.

How Long Until Markets Fully React to MP News?

Estimation results challenge the 30-minute convention. Every asset studied requires at least 40 minutes to reach price stability. As shown below in Figure 3, I find that while short-horizon assets (0--1 month) react within 40 minutes, long-term assets like 30-year Treasury futures require 50 to 60 minutes.

This suggests markets price in immediate target changes quickly but require more time to process “soft” information regarding the future policy path. This reflects a cognitive and algorithmic delay; the market must resolve disagreement about what future intentions mean for long-horizon assets.

What Happens to Monetary Surprises and Shocks?

Effects of Window Choice on Monetary Policy Surprises

Figure 7 shows that the correlation between surprises measured in 30-minute versus optimal windows declines as maturity increases. This decline is more pronounced at longer underlying maturities. For example, while surprises for near-term meetings are highly correlated (above 95%) regardless of window choice, the correlation for 15-year-ahead surprises (measured using 50-minute windows) drops by 10%. This demonstrates that as expectations about monetary policy move further into the future, the resulting interest-rate surprises become more sensitive to the choice of event window.

Effects of Window Choice on Monetary Policy Shock Impacts

Shocks derived from optimal windows exhibit larger peaks and troughs, seen below in Figure 8. In critical meetings like September 2008, the 50-minute optimal window reveals that a “central bank information shock” (CBI) was the dominant driver, whereas the 30-minute window characterises it as a standard policy shock. This suggests processing the Fed’s positive signal about economic health took significantly longer than pricing the rate action itself.

The impact of forward guidance shocks on real Treasury yields is 23 to 26 basis points larger than estimates from standard windows, shown below in Table 1. Similarly, the negative response of the S&P 500 is up to 18% stronger when the window is corrected.

Furthermore, impulse responses for CPI calculated using lag-augmented local projection (Olea and Plagborg-Møller, 2021) are nearly 10% more precise when using optimal window shocks (Figure 9 below).

Conclusion

This paper demonstrates that event window duration should be an empirical parameter rather than an ad-hoc assumption. By leveraging a neural network to approximate the underlying relationship between statement text and price discovery, I provide a methodology for identifying the optimal window that captures full market reactions.

The conventional 30-minute window is consistently insufficient, especially as the Federal Reserve relies on complex communication to shape expectations. As the literature expands to study more nuanced central bank signalling, ensuring our empirical tools accurately reflect the time markets require to process such information is essential. Adopting data-driven window lengths corrects for attenuated measurement of policy shocks and provides a more precise foundation for understanding the transmission of monetary policy to the broader macroeconomy.

About the Author

Paul L. Tran is a Ph.D. candidate at the University of Texas at Austin.

His research interests are Macroeconomics, Monetary Economics, Text Analysis, and Machine Learning. To learn more about his research, visit: https://paulletran.com/

Social Media Handles:

• LinkedIn
• Mail