Attention and Technical Analysis: State‑Dependent Predictability of CNN‑Based Stock Return Forecasts
Researcher ORCID Identifier
https://orcid.org/0009-0003-2481-5888
Graduation Year
2026
Date of Submission
12-2025
Document Type
Open Access Senior Thesis
Degree Name
Bachelor of Arts
Department
Economics
Reader 1
Professor Peter Kelly
Rights Information
2025 Crystal Y Ma
Abstract
This thesis studies how investor attention around Federal Open Market Committee (FOMC) announcements affects cross-sectional return predictability measured by a convolutional neural network (CNN) trained on price charts. I replicate Jiang, Kelly, and Xiu’s (2023) image-based CNN using 20-day OHLCV “price images,” train it on U.S. equities from 1993–2000, freeze the model, and apply it out of sample to 2001–2024 CRSP data. The CNN assigns each stock a weekly probability of outperforming which is then sorted into deciles that form high–minus–low (H–L) long–short portfolios at horizons of 1 to 10 trading days. For matched non-FOMC (control) weeks, the equal-weighted CNN signal is considerably high and is measured by the annualized H–L spread of 70.74% (22.69% value-weighted). I examine this signal within a macro-event study in order test my attention-based efficiency hypothesis. For the event study, I identify 216 scheduled FOMC announcements and match each FOMC week to five non-FOMC weeks with similar timing but at least ten trading days away from any meeting. I then compare CNN H–L spreads between FOMC and matched weeks. Around FOMC announcements, the equal-weighted 1-day spread falls to 0.10%, which is an 89% reduction relative to a matched week. I find that the event-level difference of –0.78 percentage points is highly significant. Additionally, the compression is strongest in small-cap, equal-weighted portfolios. The model shows the drop in H–L spreads is roughly 4.3 times larger for equal-weighted than value-weighted portfolios and remains negative across 1 to 10 day horizons. Overall, these results show that machine-learning-based technical predictability is strongly state dependent and largely concentrated in small-cap stocks. When macroeconomic news draws more investor attention, markets appear substantially more efficient with respect to CNN-based technical signals. This provides empirical support for limited-attention and state-dependent efficiency theories.
Recommended Citation
Ma, Crystal Y., "Attention and Technical Analysis: State‑Dependent Predictability of CNN‑Based Stock Return Forecasts" (2026). CMC Senior Theses. 4298.
https://scholarship.claremont.edu/cmc_theses/4298
Data Repository Link
https://github.com/crystalma22/cnnthesis/blob/replication-final
