HOME / RESEARCH 03 — ACADEMIC RESEARCH
Curiosity,
with a method.
Questions become research
when evidence enters the conversation.
Two studies connecting mathematical models
with financial returns and musical patterns.
01 / FINANCIAL ECONOMICS
Evaluating the
Fama–French model
↗
THE QUESTION
How do three factors explain stock returns?
ABSTRACT / SUMMARY
Using CSMAR observations from 2000–2023, this study tests the Fama–French three-factor model through regression, confidence intervals and hypothesis tests. The reported model explains 26.24% of return variation, with positive market and size coefficients and a negative value coefficient.
Asset pricing / Three-factor model / Regression
THREE FACTORS. ONE QUESTION.
MKT
Market excess return
SMB
Small minus big
HML
High minus low
MODEL STRUCTURE / CONCEPTUAL
2000–2023
Sample period
691,245
Reported observations
0.2624
Reported R²
CSMAR
Data source
METHOD
Stata regression · Confidence intervals
Coefficient & joint hypothesis tests
KEY FINDING
Positive market and size coefficients; a negative value coefficient.
Read the full paper ↗
MKT
0.0100038
Reported coefficient · Table 2
SMB
0.0088249
Reported coefficient · Table 2
HML
−0.0008635
Reported coefficient · Table 2
02 / AI × MUSIC INFORMATION RETRIEVAL
Hearing genres.
Tracing connections.
↗
TMGAT: Temporal Multi-view Graph Attention Networks
for Music Genre Classification and Evolution Mining
Peiyu Zeng & Shang Ma · Co-authored research
ABSTRACT / SUMMARY
TMGAT combines acoustic descriptors, PANNs and wav2vec 2.0 with contrastive learning and temporal graph attention to classify music and explore relationships between genres. The paper reports 82.5% accuracy on FMA-large and 85.7% in cross-corpus GTZAN testing. Its temporal analysis examines how genre patterns change within the sampled corpus.
Graph attention / Music classification
Multi-view learning / Temporal evolution
METHOD AT A GLANCE
01 / Represent
Acoustic + PANNs + wav2vec 2.0
02 / Connect
Contrastive learning + k-NN graph
03 / Model
Temporal GAT + domain adaptation
FIG. 1 / MODEL ARCHITECTURE — FROM THE PAPER
82.5%
FMA-large accuracy
25,000 tracks · 16 classes
79.6%
MTG-Jamendo accuracy
55,701 tracks · 16 classes
85.7%
GTZAN transfer accuracy
1,000 tracks · 10 classes
Results reported in the manuscript, averaged over five seeds. GTZAN is a cross-corpus test.
Which components matter?
Ablation experiments compare the full model with individual components removed.
How do patterns change?
Estimated genre proportions within FMA-large across decades; not a measure of all music.
Read the full paper ↗
METHODS & DATA
Different questions.
Deliberate methods.
Quantitative inquiry
Market data → Regression → Statistical tests
Computational inquiry
Audio representations → Temporal graphs → Classification & evolution
PAPERS & RESEARCH MATERIALS
Read the thinking behind the work.
Shang Ma RESEARCH / SELECTED PAPERS