SarthakVishnu/dissertation-dataset

Dissertation Dataset — Hybrid Topic and Domain-Adaptive Modelling for Financial Risk and Forecasting

Institution: University of Edinburgh, School of Informatics Supervisor: Prof Tiejun Ma

This is a private research dataset. All files are used exclusively within the scope of this dissertation project and must not be redistributed.

---

What This Repo Contains

This repo holds every dataset the submitted thesis depends on, plus the scripts that built them, laid out so raw sources, pipeline intermediates and final deliverables are never mixed together. See data_collection_readme.md for full step-by-step provenance of every file.

dissertation-dataset/
├── README.md                          ← this file
├── data_collection_readme.md          ← full collection/processing log
├── sp500_1A.tar.gz                    ← primary text corpus
├── sp500_union_constituents(1).csv    ← S&P 500 universe definition
├── feature_table.parquet              ← Phase 1 final table (identifiers + labels)
├── feature_table_preview.csv          ← human-readable preview of the above
│
├── raw/                                ← untouched source files
├── intermediate_data_ref_only/         ← pipeline intermediates (reference only, not final)
├── compustat/                          ← firm fundamentals
├── form4/                              ← insider trading transactions
├── calls/                              ← earnings-call transcripts
├── gdelt/                              ← NOT used in the final thesis (kept for record)
├── sec/                                ← NOT used in the final thesis (kept for record)
└── scripts/                            ← collection/processing scripts + environment files

The Five Datasets the Thesis Depends On

Per Appendix B ("Data sources") of the submitted thesis:

#	Data	Path in this repo	Coverage
1	Item 1A risk-factor text, filing dates, SIC codes	sp500_1A.tar.gz, intermediate_data_ref_only/filings_index.csv	2006–2026
2	Daily prices and returns	raw/all_sp500_prices_2000_2024_delisted_include.csv (FINSABER), raw/crsp_2025_daily.csv (CRSP)	2000–2024, 2025
3	Firm fundamentals (market cap, leverage, book-to-market, ROA)	compustat/compustat_fundamentals.csv	2006–2025
4	Insider transaction records	form4/form4_transactions.csv	2006–2026
5	Earnings-call transcripts	calls/SP500_calls_2006to2025.parquet	2006 – early 2025 (39,501 transcripts)

gdelt/ (news sentiment) and sec/ (full 10-K/10-Q text) were collected during earlier exploratory work but do not appear anywhere in the submitted methodology or results — nothing in the reproducibility path below depends on them.

---

Full File Listing

Path	Size	Description
sp500_1A.tar.gz	~115 MB (compressed)	Item 1A Risk Factor text corpus — 8,247 pickle files, one per {TICKER}_{YEAR}, covering S&P 500 firms 2006–2025. Each pickle is a plain string of the full Item 1A section.
sp500_union_constituents(1).csv	8.1 KB	S&P 500 universe — 656 unique CIK/ticker pairs (historical constituents).
feature_table.parquet	~0.2 MB	Phase 1 master table — 8,105 rows, one per filing. Columns: ticker, cik, permno, sic, fiscal_year, filing_date, report_date, lagged_vol_30d, fwd_vol_30d, plus placeholder embedding/topic_vector columns filled later in the pipeline.
feature_table_preview.csv	~1.0 MB	Human-readable CSV version, without placeholder columns.
raw/all_sp500_prices_2000_2024_delisted_include.csv	265 MB	Daily prices 2000–2024, all S&P 500 firms including delisted (FINSABER, public).
raw/crsp_2025_daily.csv	9.99 MB	Daily prices for 2025 (WRDS CRSP, institutional licence).
raw/crsp_dsenames.csv	301 KB	CRSP security names history — one row per name period per PERMNO.
raw/ccm_linking_table.csv	2.21 MB	Raw CIK → GVKEY → PERMNO/PERMCO linking table (WRDS CCM).
raw/ticker_cik_mapping(1).csv	394 KB	Auxiliary ticker↔CIK cross-reference.
intermediate_data_ref_only/permno_linkage.csv	36.9 KB	Cleaned CIK → PERMNO mapping (651/656 matched).
intermediate_data_ref_only/filings_index.csv	961 KB	EDGAR filing dates + SIC codes for all 656 CIKs.
intermediate_data_ref_only/volatility_labels.csv	780 KB	30-day annualised volatility windows per filing.
intermediate_data_ref_only/permnos_for_wrds.txt	4.53 KB	PERMNO list used to query WRDS for crsp_2025_daily.csv.
compustat/compustat_fundamentals.csv	1.16 MB	Annual accounting/market fundamentals, 480 firms, 2006–2025 (WRDS Compustat).
form4/form4_transactions.csv	263 MB	1,670,042 non-derivative insider transactions, 2006–2026.
form4/form4_preview.csv	2.28 KB	Head sample of the above for quick inspection.
calls/SP500_calls_2006to2025.parquet	1.07 GB	Consolidated earnings-call transcript corpus, 2006–2025.
calls/firm/json/**	~19 MB	Original 299-file/172-ticker JSON dump (2025–2026), superseded by the parquet above but kept for provenance.
gdelt/gdelt_gkg_daily.parquet	6.7 MB	(unused in final thesis) Daily news-sentiment features, 543 PERMNOs, 2025–2026.
sec/sec_filings_index.csv	352 KB	(unused in final thesis) Index of full 10-K/10-Q filing text (text itself not uploaded).
scripts/*	39 KB+	Collection/processing scripts for the datasets above; see data_collection_readme.md § Scripts Reference for the full list and what's included.

---

How to Download on the Cluster

# One-time login
huggingface-cli login

# Download all files
huggingface-cli download SarthakVishnu/dissertation-dataset \
    --repo-type dataset \
    --local-dir ~/dissertation/datasets/

# Unzip the Item 1A corpus
cd ~/dissertation/datasets/
tar -xzf sp500_1A.tar.gz && rm sp500_1A.tar.gz

To skip the unused pilots (gdelt/, sec/) and the large calls/firm/json/ snapshot, use --include/--exclude patterns with huggingface-cli download instead of a full mirror.

---

How Each Dataset Is Used in the Thesis

Dataset	Used for
Item 1A text (sp500_1A.tar.gz)	Domain-adaptive pretraining (DAPT), three-view contrastive fine-tuning, BERTopic, full-text TF-IDF baseline
Structured features (prices, Compustat)	Realised-volatility features (HAR-RV), market beta, size, leverage, book-to-market, ROA, momentum, illiquidity — the structured baseline the text representations are compared against
Form 4 insider trades	Eight insider-trading features (Appendix D) — where in the cross-section the text increment is earned (§4.5)
Earnings-call transcripts	Call-tone features contrasted against 10-K text as a second text modality (§4.4)
feature_table.parquet	Reusable deliverable joining identifiers, embeddings, topic vectors and volatility labels — one row per filing

---

Temporal Split

Split	Criterion	Approx. Filings	Role
Train	filing_date < 2025-01-01	~7,700	DAPT, contrastive FT, BERTopic
Val	2025-01-01 ≤ filing_date < 2026-01-01	~280	Perplexity checkpointing, FinMTEB eval
Test	filing_date ≥ 2026-01-01	~125	Held-out volatility forecasting

---

Licence & Ethics

• Item 1A text, EDGAR-derived data and Form 4 insider-transaction data are from publicly available SEC sources. No licence restrictions.
• FINSABER prices are sourced from the public HuggingFace dataset waylonli/FINSABER-reproduce. No redistribution restrictions for research use.
• WRDS/CRSP-derived data (raw/crsp_2025_daily.csv, raw/crsp_dsenames.csv, raw/ccm_linking_table.csv) and WRDS/Compustat data (compustat/compustat_fundamentals.csv) are used under the University of Edinburgh's institutional licence and must not be redistributed externally.
• Earnings-call transcripts were accessed and processed via the FinStreamAI data infrastructure (University of Edinburgh), originally sourced from API Ninjas.
• No personal data or human subjects are involved.