Training-Period Association, Temporal Non-Replication and Economic Limits
Peter Prendergast · MSc Computational Finance · University College London · 2026
Does the share of news stories classified as negative by FinBERT explain stock returns after controlling for average sentiment and news volume?
The negative association found in 2011–2019 did not replicate in 2020–2023. Timing diagnostics and trading costs further limit its usefulness as a predictive signal. The study covers 715,546 news-bearing firm-days across 570 priced firms in FNSPID, with separate LSEG/Reuters transfer checks.
Dissertation PDF · Evidence map · Reproduction guide · Documentation
Annual estimates from the dissertation. Horizontal lines show the training- and testing-period means. The figure is reproduced from the committed aggregate evidence.
| Period or contrast | Coefficient | 95% HAC interval |
|---|---|---|
| Training, 2011–2019 | −0.00831 | [−0.01410, −0.00252] |
| Frozen testing, 2020–2023 | +0.00583 | [−0.00426, +0.01593] |
| Testing minus training | +0.01414 | [+0.00250, +0.02579] |
These are daily cross-sectional rank-regression coefficients, with five-lag HAC intervals. They are not percentage returns. The predeclared contrast supports a difference across periods, while limited testing-period power leaves uncertainty about small effects.
An association also appears before the assigned trading window. Row-level news availability times are missing, so the data cannot establish that the news could have been traded at the assumed entry time. The best break-even cost among the tested daily rules was approximately 0.625 bps per side, against 10 bps charged.
The testing period had informed some earlier design choices, which limits a confirmatory interpretation. The research design records the chronology, multiplicity corrections and limitations alongside the null and adverse results.
From the repository root, with macOS on Apple silicon, Python 3.12, uv and Make:
make setup
make validateThis checks the final package's file hashes, manuscript references, prose, analytical tests and lint, then regenerates 30 figure and table artifacts for byte-for-byte comparison. No model calls or licensed inputs are needed. The frozen testing-period experiment remains sealed.
make manuscript builds the PDF if latexmk is installed. The committed portfolio PDF has the assessment identifier removed from its cover, with the research content unchanged. See the reproduction guide for platform limits, direct commands and the authorised inputs required for a full computational replay.
The final analysis lives in submission/news-sentiment-beyond-mean/, with 22 selected notebooks, statistical helpers, frozen specifications and aggregate results. Its evidence map traces each claim to the notebook and files that produced it.
The surrounding Python package contains the earlier benchmark and collection tools:
| Area | Source | Tests |
|---|---|---|
| Atomic news-corpus publication | Artifact writer | Tavily, NewsAPI |
| Trading-window alignment and funded portfolios | Trading pipeline, portfolio accounting | Pipeline, portfolio |
| Classification metrics and bootstrap comparisons | Metrics | Statistics |
| Artifact hashes and run provenance | Artifact I/O, runtime metadata | Metadata, project checks |
| Command-line and terminal interfaces | CLI, TUI | CLI, TUI |
These tools have a separate locked environment:
make benchmark-setup
make benchmark-checkThe check validates the 5,947-row Financial PhraseBank/FiQA benchmark, runs the root test suite, and checks Ruff and mypy. To try the application, start with the benchmark guide. The architecture guide explains the modules and data flow.
final_experiments/, scripts/, configs/ and reports/ retain the earlier research. Paths remain stable so old results can still be traced. The archive map explains superseded designs and notebook-number overlaps.
Licensed news text, raw model responses, credentials and intermediate panels stay outside Git. This repository contains material with mixed ownership and no blanket open-source licence. See data and rights before reuse and CITATION.cff for the dissertation citation.
Contributing · Security · CI