EarningsOrderflowStudy

EarningsOrderflowStudy: A Minute-Level Event Study of Microsoft Earnings Announcements.

This project characterizes how MSFT's price, volume, and orderflow proxies respond in the minutes around the last 5 quarterly earnings announcements. It bridges the HawkesLOB framework conceptually to a real-world event-study application using Polygon minute bars and SEC 8-K filings.

Averaged Event Signature (N=5)

The plots below show the canonical "event signature" across the sample. Note the massive volume spike at $\tau=0$ and the subsequent decay in realized volatility.

Findings & Data Summary

Event Panel Results

The following table summarizes the 5 most recent MSFT earnings events as of the study date.

Date	Net Sentiment	EPS Surprise	Return at +30m
2026-04-29	-0.62%	+5.17%	-1.53%
2026-01-28	-0.61%	+5.61%	-3.46%
2025-10-29	-0.61%	+12.84%	-2.17%
2025-07-30	-0.80%	+7.99%	+7.66%
2025-04-30	-0.95%	+7.45%	+7.10%

Note: Net Sentiment is calculated using the official, full Loughran-McDonald (L-McD) Master Dictionary. Net Sentiment = (Positive Words - Negative Words) / Total Words. Surprise is standardized: (Actual - Estimate) / |Estimate|.

Analysis: Sentiment vs. Surprise

We analyzed how news features predict the jump magnitude at $\tau=30$ minutes from the press release. In this small sample, the standardized EPS surprise and L-McD sentiment features show starkly different relationships with the initial price reaction.

Individual Event Paths

While the averaged signature highlights the general trend, individual events show significant variance in both magnitude and drift. The per-event overlay is particularly valuable as it reveals a distinct bipolar response pattern rather than a uniform drift: two strong gainers (H1 2025) and three post-announcement fades (Q3 2025 onwards).

Specifically, the 2026-04-29 event shows a classic "up-then-down" intraday trajectory. In microstructure terms, this represents a mechanical price reaction to the positive headline EPS beat ($4.27 actual vs. $4.06 consensus) which was rapidly faded as the market absorbed more detailed guidance and segment numbers crossing the wire minutes later.

Key Conclusions

1. The "Beat" is Completely Priced In: Despite all 5 quarters exhibiting positive EPS surprises, MSFT saw negative 30-minute responses in 3 out of 5 events. Standardized EPS surprise has 0.00% correlation ($R^2 = 0.0000$) with the post-earnings stock jump in this sample! This suggests that the headline beat is fully anticipated by the market and that other forward-looking factors (like revenue segments or management guidance) completely dominate the price reaction.
2. Spurious Sentiment Correlation: While the official L-McD dictionary shows an apparent extremely strong negative correlation ($R^2 = 0.8503$, $r = -0.9221$) between net sentiment and the cumulative return at $\tau = 30$, this relationship is almost certainly spurious. Both variables are perfectly confounded with time period: older events (H1 2025) had both more-negative sentiment and positive returns, while recent events (Q3 2025 onwards) had less-negative sentiment and negative returns. Furthermore, the absolute sentiment difference is tiny (~2 net negative words per 1000 words), well within natural press-release stylistic variation. A defensible analysis would require $N \ge 30$ events spanning multiple market regimes, using sentiment as one of several controls.
3. Microstructure Alignment: The volume spike now peaks precisely at $\tau = 0$, validating our uniform $-7$ minutes shift which successfully calibrated the SEC 8-K filing timestamp to the actual BusinessWire release time. Realized volatility peaks at $\tau = 0$ and decays steadily back to baseline over ~45 minutes.
4. Orderflow Pressure: The CVD (Cumulative Volume Delta) proxy shows that the direction of the initial jump is strongly correlated with the net signed volume in the first 5 minutes, confirming that orderflow leads the settling price.
5. Pre-Event Drift & Information Leakage: We observe that cumulative returns rise from 0 to ~0.008 between $\tau = -20$ and $\tau = 0$, accompanied by a gradual build in volume during the same window. This is consistent with microstructure literature on pre-announcement informed trading (e.g., Lee, Ready, and Seguin 1992), though a much larger sample is required to test for statistical significance.

---

Features

• Minute-level Resolution: Characterizing microstructure response in extended-hours sessions.
• Event Study Methodology: Averaged response curves (MacKinlay 1997) with standard error bands.
• News Features: Loughran-McDonald sentiment analysis on SEC 8-K press releases.
• Orderflow Proxies: Tick-rule signed volume, CVD proxy, and Corwin-Schultz spread estimates.

Data Sources

• Polygon.io: Minute-bar OHLCV with extended hours.
• SEC EDGAR: 8-K filings and Exhibit 99.1.
• yfinance: EPS actuals and consensus estimates.
• Loughran-McDonald: Financial sentiment master dictionary (Baseline version included).

Reproducibility Guide

Follow these steps to reproduce the study from scratch.

1. Environment Setup

Clone the repository and create a virtual environment:

git clone https://github.com/yourusername/EarningsOrderflowStudy.git
cd EarningsOrderflowStudy
python3 -m venv venv
source venv/bin/activate  # On Windows use `venv\Scripts\activate`
pip install -r requirements.txt

2. Configure API Keys

The study requires a Polygon.io free-tier API key for minute-level bars.

1. Create a .env file in the root directory.
2. Add your key:

   POLYGON_API_KEY=your_actual_key_here

3. Data Acquisition

Fetch the last 5 earnings dates, SEC 8-K filings, and Polygon minute bars. This script includes a fallback to generate a baseline sentiment dictionary if the official download is unavailable.

export PYTHONPATH=$PYTHONPATH:.
python3 scripts/fetch_all.py

4. Run Analysis

You can generate the plots and summary tables using the provided Jupyter notebook:

1. Launch Jupyter: jupyter notebook
2. Open notebooks/01_analysis.ipynb and run all cells.

Alternatively, the analysis logic is encapsulated in src/ and can be imported into any research script.

License

Apache 2.0