Audio Alignment Method Comparison (earnings calls, 943 files)
| Method |
Median Error |
Time / Call |
Notes |
|
| Fuzzy / rapidfuzz |
<15 s |
~4 min |
Clip-based word refinement |
SELECTED |
| WhisperX |
<15 s |
~12 min |
3× slower; same accuracy |
REJECTED |
| pyannote |
732 s |
~20 min |
Turn-level diarisation mismatch |
UNUSABLE |
pyannote's 732s error is a structural failure: it cannot recover word-level timestamps from PDF transcripts where speaker turns don't align with audio segments.
Capital IQ PDF Parser Comparison (n=5 ground-truth PDFs)
| Parser |
Company Recall |
Analyst Contamination |
Runtime |
Deps |
|
| pdfplumber |
1.00 |
0.00 |
2.72 s |
Minimal |
SELECTED |
| MarkItDown |
1.00 |
0.00 |
2.88 s |
Heavier |
RUNNER-UP |
| pymupdf4llm |
0.80 |
0.00 |
6.31 s |
Heavier |
REJECTED |
pdfplumber selected: perfect recall, zero contamination, fast runtime, minimal dependencies. Known failure mode: 2-column participant pages (~20% of PDFs) — handled via bounding-box extraction in the production parser.
+0.95
Net Sharpe (OOS) · CI [−0.19, +2.13]
+14.3%
Annualised net return
+48.9%
Cumulative return · 751 OOS days
Walk-Forward Fold Results — Condition 1 (3 folds, 2023–2025)
| Fold |
Test Year |
HL Selected |
Net Sharpe |
95% CI |
Max DD |
Hit Rate |
| 1 |
2023 |
7d |
+0.39 |
[−1.66, +2.40] |
−12.6% |
51% |
| 2 |
2024 |
7d |
+1.57 |
[−0.40, +3.47] |
−10.4% |
49% |
| 3 |
2025 |
7d |
+0.88 |
[−1.05, +2.70] |
−14.0% |
50% |
| Combined OOS (751 days) |
+0.95 |
[−0.19, +2.13] |
−14.0% |
50% |
Signal: FinBERT earnings sentiment × exp(−ln(2)/7 × calendar_days_since_last_call). Long top 10 / short bottom 10 of 49 S&P 100 tickers. 5bp one-way TC. Daily rebalancing. 7-day half-life wins all 3 validation folds independently — no cherry-picking.
Fama-French 3-Factor Attribution
| Factor |
Loading |
Interpretation |
| Alpha (annualised) |
+10.1%/yr |
Unexplained excess return |
| Market (β) |
−0.037 |
Market-neutral confirmed |
| SMB (size) |
Small |
No systematic size tilt |
| HML (value) |
Small |
No systematic value tilt |
Academic Benchmark Comparison
| Study |
Signal |
OOS Sharpe |
| This study (C1) |
FinBERT, 7d HL, daily |
+0.95 |
| Price et al. 2012 |
LM tone, quarterly |
+0.6–0.8 |
| PEAD.txt (Fed, 2021) |
Text regression, LCap |
~+0.7 |
Condition 1 is a positive-but-noisy baseline. CI [−0.19, +2.13] barely excludes zero on 751 days of data — this is unavoidable with short OOS windows. The research question is whether Conditions 2/3/4 (earnings audio, podcast sentiment, virality weighting) add materially on top. C2 can run immediately; C3/C4 blocked on Whisper transcription of 200 podcast MP3s.