FOMC-event-study

Does the US large-cap equity market price FOMC rate decisions efficiently? Event study of 30 firms across seven sectors × 9 FOMC rate-change announcements from March 2022 through December 2025 (270 firm-event observations), with a Fama–French 3-factor expected-return benchmark, a hike-vs-cut split, sector-level t-tests, and an event-clustered bootstrap. Pooled CAAR is statistically zero in every test window. The cross-section is not — banks gain a significant +1.23 % in [−1, +1] (t = +3.41, p = 0.001) and utilities lose a significant −0.83 % (t = −3.28, p = 0.002). Three implementations: Python primary, R port, SAS port.

TL;DR

	Estimate	t	p
Pooled CAR, [−1, +1] window	+0.086 %	+0.523	0.601
Pooled CAR, [0, +1] window	+0.143 %	+1.097	0.273
Pooled CAR, [0, +5] window	+0.003 %	+0.013	0.989
Banks, [−1, +1] window	+1.233 %	+3.41	0.001
Utilities, [−1, +1] window	−0.826 %	−3.28	0.002
Bootstrap 95 % CI on [0, +5]	[−0.55 %, +0.42 %]

Pooled CAAR drifts to roughly −0.85 % by the announcement day, then sits there. Almost all of the drift is realised between t = −5 and t = −2, before the announcement; from t = −1 onward the trajectory is flat. The pattern is consistent with semi-strong-form market efficiency at the aggregate level. But the pooled null hides a sharp cross-sectional pattern: banks and utilities both reject zero at the 1 % level in the immediate window, and the cross-sectional spread of mean CAR is an order of magnitude larger than the pooled mean. A market-wide event study tests a particular kind of efficiency; it is silent on cross-sectional predictability, and the FOMC sample examined here is a clean illustration of the distinction.

The result survives:

• splitting the sample by direction (hikes vs cuts) — neither subsample rejects zero
• an event-clustered bootstrap (2 000 replications, resampling whole events with replacement) — bootstrap SE is 18 % wider than the naive i.i.d. SE, but the 95 % CI still solidly contains zero

Quickstart

git clone https://github.com/Guannings/fomc-event-study.git
cd fomc-event-study
pip install -r requirements.txt
python event_study.py

Runtime is about 2 minutes (most of it is the 2 000-replication bootstrap on the [0, +5] window). All outputs land in ./results/, overwriting the files already shipped with the repo.

Reproducibility is exact up to floating-point precision: bootstrap seed is fixed (np.random.seed(42)), the FOMC event dates are hardcoded, the firm list is hardcoded, and the data window is closed on both ends.

What's in this repo

File	Purpose
event_study.py	Primary Python implementation. Pulls prices, runs the 270 FF3 regressions, computes CAR/CAAR, runs t-tests, executes the three robustness blocks, generates the plots.
event_study.R	R port with the same statistical output. Uses quantmod for prices, base lm.fit() for regressions, data.table for the panel, ggplot2 for plots.
event_study.sas	SAS port. SAS has no native Yahoo/Ken French API, so the workflow is two-step: dump prices.csv and ff3.csv from Python or R first, then SAS reads them. The estimation uses PROC REG inside a macro loop over 270 (firm × event) pairs.
paper.pdf	Full 18-page writeup. Methodology, tables of results, robustness section, discussion, conclusion, references.
results/	Generated outputs (see table below).
requirements.txt	Python deps. Pinned to working versions.

results/ contents:

File	Contents
betas.csv	Per (firm, event) α + 3 β estimates + R². 270 rows.
aar_caar_by_t.csv	Average abnormal return and cumulative average abnormal return by event-time t (11 rows from −5 to +5).
car_per_firm_event.csv	Firm-event CAR for each of the three test windows. 270 rows × 3 windows.
test_results.csv	Pooled one-sample t-tests for the three headline windows.
split_hike_cut.csv	Pooled t-tests separately for the four hikes and the five cuts.
sector_ttests.csv	Sector-level one-sample t-tests by window.
bootstrap_ci_0_5.csv	Event-clustered bootstrap 95 % CI for the [0, +5] window.
caar_plot.png	Main CAAR-trajectory plot (the one in the paper).
caar_by_sector.png	Sector breakdown of CAAR over the event window.

Method

Sample

Thirty US-listed, large-capitalisation firms hand-picked to span the cross-section of theoretical interest-rate sensitivity:

Sector (n)	Hypothesis on AR following a hike	Tickers
Banks (6)	AR > 0 (net interest margin)	JPM, BAC, WFC, C, GS, MS
REITs (6)	AR < 0 (bond proxy)	O, SPG, PLD, EQIX, AMT, WELL
Utilities (5)	AR < 0 (bond proxy)	NEE, DUK, SO, AEP, EXC
Tech (6)	AR < 0 (long duration)	AAPL, MSFT, GOOGL, AMZN, NVDA, META
Consumer Disc. (3)	AR < 0	HD, NKE, SBUX
Industrials (2)	AR ≲ 0	CAT, DE
Energy (2)	control (weak link)	XOM, CVX

For rate cuts the predicted signs reverse.

Events

Nine FOMC target-rate change announcements, verified from federalreserve.gov/monetarypolicy/openmarket.htm:

Date (t = 0)	Action	Notes
2022-03-17	+25 bp	First hike of the post-COVID cycle
2022-06-16	+75 bp	First +75 bp move since 1994
2022-09-22	+75 bp	Hawkish surprise; hot prior-week CPI
2023-03-23	+25 bp	Hike during the SVB / regional-bank stress
2024-09-19	−50 bp	First cut of the cycle; market pricing split
2024-12-19	−25 bp	"Hawkish cut" (dot-plot shock)
2025-09-18	−25 bp	Opening cut of the 2025 easing cadence
2025-10-30	−25 bp	Continuation cut
2025-12-11	−25 bp	Final cut before extended pause

Four hikes, five cuts. The sample spans the full hiking cycle of 2022–2023, the September 2024 pivot, and the 2025 easing cadence.

Factor model and event windows

For each (firm i, event e) pair, factor loadings are estimated by OLS on the trading-day window τ ∈ [−250, −30] relative to the event day t = 0:

$$R_{i,\tau} - RF_\tau = \alpha_{i,e} + \beta^{MKT}_{i,e},(MKT - RF)_\tau + \beta^{SMB}_{i,e},SMB_\tau + \beta^{HML}_{i,e},HML_\tau + \varepsilon_{i,\tau}$$

That gives roughly one calendar year of returns, ending one calendar month before the event so the window is not contaminated by pre-event drift or leakage. With 30 firms × 9 events = 270 (firm, event) pairs, the procedure runs 270 separate OLS regressions.

Expected returns on the event window [−5, +5] are then constructed by applying the estimated betas to the event-window factor realisations:

$$\widehat{E[R_{i,t}]} = RF_t + \hat\alpha_{i,e} + \hat\beta^{MKT}_{i,e},(MKT - RF)_t + \hat\beta^{SMB}_{i,e},SMB_t + \hat\beta^{HML}_{i,e},HML_t$$

Abnormal returns and cumulative abnormal returns follow as AR_i,t = R_i,t − Ê[R_i,t] and CAR_i(τ₁, τ₂) = Σ AR_i,t over t ∈ [τ₁, τ₂].

Test windows

Three windows are reported, each satisfying the requirement that it contain the event day plus a post-event period: [−1, +1] for the immediate reaction, [0, +1] for the announcement day plus one, and [0, +5] for the delayed reaction or drift. A Bonferroni correction across the three windows would set the per-test level at α = 1.67 %; the qualitative conclusions do not depend on which level is used.

Data sources

• Prices: Yahoo Finance via yfinance. Daily adjusted close from 2021-01-01 through 2026-06-13. The early start gives ~250 trading days of history before the first event for the estimation window.
• Factors: Kenneth French's Data Library at Dartmouth Tuck. Daily research factors (Mkt-RF, SMB, HML, RF). The factor series is published in percent and is divided by 100 before merging with returns — forgetting this conversion is a frequent source of error in event-study replications.
• FOMC dates: Federal Reserve Board's Open Market Operations history.

Results

Aggregate CAAR trajectory

Median R² across the 270 regressions is 0.430 (range ~0.10 for energy names to ~0.80 for the largest banks), comfortably within the typical range for FF3 on US large-cap equities.

The pooled CAAR drifts from ≈ −0.14 % at t = −5 down to ≈ −0.85 % at t = 0, with the bulk of the decline concentrated in the [−3, −2] window (combined AAR ≈ −0.56 %). From t = −1 onward the trajectory is flat. None of the three test windows rejects the null:

Window     Mean CAR (%)   Std (%)    t-stat     p      N
[-1, +1]      +0.086       2.71      +0.523   0.601    270
[ 0, +1]      +0.143       2.14      +1.097   0.273    270
[ 0, +5]      +0.003       3.51      +0.013   0.989    270

Sector-level t-tests

The pooled null hides the more interesting result. Sector-mean CARs:

                       [-1, +1]            [0, +1]             [0, +5]
                  Mean    t      p     Mean    t      p     Mean    t      p
Banks         +1.233***  +3.41 0.001  +0.879***  +3.05 0.004  +0.722*   +1.98 0.053
Industrials   +1.462*    +1.83 0.085  +0.509     +0.93 0.367  +0.944    +0.94 0.362
Tech          +0.348     +0.80 0.425  -0.064     -0.18 0.858  -0.648    -1.16 0.250
Consumer D.   +0.039     +0.07 0.943  +0.141     +0.36 0.723  +0.199    +0.28 0.784
REITs         -0.552*    -1.97 0.054  +0.164     +0.64 0.524  +0.113    +0.24 0.812
Utilities     -0.826***  -3.28 0.002  -0.162     -0.78 0.439  -0.465    -0.92 0.360
Energy        -1.249*    -1.98 0.064  -1.110*    -2.01 0.061  -0.597    -0.83 0.419

Stars: *** p < 0.01, ** p < 0.05, * p < 0.10.

The two sectors at the extremes of the rate-sensitivity hypothesis both reject the null at the 1 % level in the [−1, +1] window, and both survive any reasonable multiple-comparison correction. At longer horizons the sector effects attenuate as the differential reaction is absorbed back into the index.

Hike vs cut subsamples

The result survives splitting by direction. Neither the four hikes nor the five cuts rejects the null in any test window:

Subsample           Window     Mean CAR (%)   t      p      N
Hikes (N_e = 4)     [-1, +1]      +0.178    +0.72  0.474   120
                    [0, +1]       +0.078    +0.40  0.692   120
                    [0, +5]       -0.280    -0.77  0.443   120
Cuts  (N_e = 5)     [-1, +1]      +0.013    +0.06  0.953   150
                    [0, +1]       +0.194    +1.12  0.265   150
                    [0, +5]       +0.229    +0.91  0.365   150

The hike subsample's [0, +5] is mildly negative (−0.28 %) and the cut subsample's is mildly positive (+0.23 %), directionally consistent with the theoretical signs being opposite, but both are statistically zero.

Bootstrap-by-event robustness

The t-tests above assume independent firm-event observations, but firms on the same announcement day share a common shock. The [0, +5] window is re-tested using a non-parametric bootstrap that resamples whole events with replacement (2 000 replications):

Point estimate:           +0.003 %
Bootstrap 95 % CI:        [-0.55 %, +0.42 %]
Bootstrap SE (clustered):  0.253 %
Naive (i.i.d.) SE:         0.214 %
SE inflation factor:       1.18 x

The event-clustered SE is 18 % larger than the naive SE, but the confidence interval still solidly contains zero. The non-finding is not an artefact of ignoring within-event cross-sectional correlation.

Sector breakdown plot

Banks and Industrials drift upward; Utilities, REITs, and Energy drift downward; Technology and Consumer Discretionary are approximately flat. The spread between the most positive and most negative sector means in the [0, +5] window is about 1.5 percentage points, an order of magnitude larger than the pooled mean.

Sample output

$ python event_study.py
Downloading Fama-French factors from Ken French Data Library...
  Loaded FF3: 1926-07-01 -> 2026-04-30 (26233 rows)
Downloading 30 tickers from Yahoo Finance...
  Loaded prices: 2021-01-04 -> 2026-06-12 (1367 rows, 30 tickers)

Running 270 FF3 estimations + event-window AR computations...
  Successful regressions: 270 / 270 (skipped 0)
  Median R^2: 0.430
  Pooled firm-events used in AAR/CAAR: 270

CAAR trajectory (in percent):
       AAR    CAAR
t
-5 -0.1403 -0.1403
-4 -0.0964 -0.2367
-3 -0.2358 -0.4725
-2 -0.3218 -0.7943
-1 -0.0563 -0.8507
 0 -0.0139 -0.8646
 1  0.1565 -0.7080
 2 -0.1523 -0.8603
 3  0.0702 -0.7902
 4  0.0901 -0.7001
 5 -0.1477 -0.8478

Headline t-tests:
  Window [-1, +1]:  mean CAR = +0.0863%  t = +0.523  p = 0.6014  N = 270
  Window [+0, +1]:  mean CAR = +0.1427%  t = +1.097  p = 0.2735  N = 270
  Window [+0, +5]:  mean CAR = +0.0028%  t = +0.013  p = 0.9894  N = 270

Hike vs cut split:
  hike  [-1, +1]:  mean CAR = +0.1780%  t = +0.719  p = 0.4736  N = 120 (events = 4)
  hike  [+0, +1]:  mean CAR = +0.0781%  t = +0.397  p = 0.6921  N = 120 (events = 4)
  hike  [+0, +5]:  mean CAR = -0.2797%  t = -0.770  p = 0.4427  N = 120 (events = 4)
  cut   [-1, +1]:  mean CAR = +0.0130%  t = +0.059  p = 0.9534  N = 150 (events = 5)
  cut   [+0, +1]:  mean CAR = +0.1943%  t = +1.119  p = 0.2650  N = 150 (events = 5)
  cut   [+0, +5]:  mean CAR = +0.2289%  t = +0.908  p = 0.3652  N = 150 (events = 5)

Sector-level t-tests:
  [-1,+1] Industrials        mean = +1.462%  t = +1.826  p = 0.0854  N = 18 *
  [-1,+1] Banks              mean = +1.233%  t = +3.411  p = 0.0012  N = 54 ***
  [-1,+1] Tech               mean = +0.348%  t = +0.804  p = 0.4248  N = 54
  [-1,+1] Consumer Disc.     mean = +0.039%  t = +0.073  p = 0.9425  N = 27
  [-1,+1] REITs              mean = -0.552%  t = -1.970  p = 0.0541  N = 54 *
  [-1,+1] Utilities          mean = -0.826%  t = -3.278  p = 0.0020  N = 45 ***
  [-1,+1] Energy             mean = -1.249%  t = -1.981  p = 0.0640  N = 18 *

Bootstrap-by-event CI for [0, +5] CAR (2000 replications):
  Point estimate:              +0.0028%
  Bootstrap 95% CI:            [-0.5471%, +0.4228%]
  Bootstrap (clustered) SE:    0.2525%
  Naive (i.i.d.) SE:           0.2138%
  Inflation factor:            1.18x

Done. All outputs in /…/results/

Multi-language ports

Python is the primary implementation. R and SAS ports are byte-equivalent on the numbers up to floating-point precision (and bootstrap noise, since the bootstrap RNG is not portable across languages).

R

Uses quantmod for Yahoo Finance prices, base lm.fit() for the 270 regressions, data.table for the panel manipulation, and ggplot2 for the plots.

Rscript event_study.R

Install dependencies once with install.packages(c("quantmod", "data.table", "ggplot2", "scales")).

SAS

SAS has no native Yahoo Finance or Ken French API, so the workflow is two-step: dump prices.csv and ff3.csv from the Python run first, then run the SAS script. Estimation uses PROC REG inside a macro loop over the 270 (firm × event) pairs; t-tests use PROC TTEST; plots use PROC SGPLOT.

%let work_dir = /path/to/fomc-event-study;
%include "&work_dir./event_study.sas";

Edit the %let work_dir = ... line at the top of event_study.sas to point at the repo root before running.

Caveats and known limitations

These are documented as a longer discussion in §5 of paper.pdf; the short version:

• Sample is nine events. A minimum-detectable-effect calculation at 80 % power and the observed cross-sectional standard deviations puts the pooled tests' detection threshold at roughly 40 to 60 basis points, depending on the window. The interpretation that prices fully impound the policy news is an inference from non-rejection, not from a tight bound around zero. A larger event sample would tighten the bound.
• Surprise component not extracted. All rate changes are treated symmetrically. Following Kuttner (2001), a Fed-funds-futures-derived surprise measure would let the specification regress AR on the policy surprise itself — a meaningfully cleaner test of informational efficiency.
• Market-wide event caveat. The market factor MKT − RF reacts to the FOMC announcement itself, so part of the systematic Fed-news impact is absorbed into the regression rather than surviving into AR. What's left in AR is the differential reaction relative to the FF3 benchmark — exactly the right quantity for the cross-sectional analysis, but the pooled CAAR should not be read as the absolute equity-market response to Fed news.
• Energy is reframed as a confound, not a control. The energy reaction is strongly negative across most windows. The likely explanation is that the pooled sample includes dovish moves in 2024 and 2025 that coincided with commodity-price weakness for the specific energy names — not that energy is rate-sensitive.
• Survivorship bias. All 30 firms were continuously listed through 2021–2026 by construction. Going back to pre-2008 would make this worse (META was 2012, GOOG was 2004); the 2022–2025 cycle is the longest window where the current sample composition is consistent throughout.

Future work

• Hike-vs-cut analysis is structurally limited by the four-hike sample. Extending the window back through the 2015–2018 Yellen cycle and the 2019 cuts would put the directional comparison on a real footing.
• Kuttner-style surprise decomposition using Fed-funds-futures.
• High-frequency intraday version: 30-minute equity reactions to the 14:00 ET statement release, which would identify the immediate market-wide response without the differential-reaction absorption.

License

MIT.

====================================================================================

Disclaimer and Terms of Use

1. Educational Purpose Only

This software is for educational and research purposes only and was developed as a final project for an Investments course by PARVAUX, a Public Finance and Economics double major at National Chengchi University (NCCU). It is not intended to be a source of financial advice, and the author is not a registered financial advisor or licensed securities professional. The event-study methodology implemented here — Fama–French 3-factor expected returns, OLS estimation windows, cumulative abnormal returns, one-sample t-tests, sector splits, and event-clustered bootstrapping — is a demonstration of standard empirical asset-pricing techniques and should not be construed as a recommendation to buy, sell, hold, or short any specific security or sector.

2. No Financial Advice

Nothing in this repository constitutes professional financial, legal, or tax advice. Investment decisions should be made based on your own research and consultation with a qualified financial professional in your jurisdiction. The empirical findings reported here — that the pooled CAAR around FOMC rate-change announcements is statistically zero, that the cross-section is not, that banks gain on hikes and utilities lose — are properties of a specific 2022–2025 sample under a specific factor-model specification and do not generalise without further work.

3. Statistical Caveats

• A non-rejection of the null is not evidence that the null is true. The pooled tests have limited power against effects smaller than ~40–60 bp.
• Cross-sectional t-tests assume independent observations across firms and events. Firms on the same announcement day share a common shock; the event-clustered bootstrap reported here addresses this for the [0, +5] window but not for the others.
• The Fama–French three-factor model captures market, size, and value exposures but not momentum, profitability, investment, or liquidity. Any systematic factor exposure not in the model contaminates the abnormal returns.
• The FOMC sample combines hikes and cuts. The hike-vs-cut split confirms the qualitative conclusion in this sample, but cycle composition matters for any claim about a "typical" FOMC reaction.

4. Data Provider Reliability

The author has no affiliation with Yahoo Finance, Kenneth French, the Federal Reserve Board, or any other data provider cited above and assumes no responsibility for outages, incorrect data, API breaking changes, or rate-limit denials originating from those services. The yfinance library is a third-party wrapper around Yahoo Finance's unofficial endpoints and may break at any time; Kenneth French's research factors are updated on a schedule the author does not control.

5. "AS-IS" SOFTWARE WARRANTY

THIS SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, AND NON-INFRINGEMENT. IN NO EVENT SHALL THE AUTHOR OR COPYRIGHT HOLDER BE LIABLE FOR ANY CLAIM, DAMAGES, OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT, OR OTHERWISE, ARISING FROM, OUT OF, OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

BY USING THIS SOFTWARE, YOU AGREE TO ASSUME ALL RISKS ASSOCIATED WITH YOUR INVESTMENT, RESEARCH, OR REPLICATION DECISIONS, RELEASING THE AUTHOR (PARVAUX) FROM ANY LIABILITY REGARDING YOUR FINANCIAL OUTCOMES OR ACADEMIC USES OF THE METHODS, FINDINGS, OR CODE PRESENTED.

====================================================================================