Sashank-810/crisisnet-dataset

CrisisNet — Corporate Default Risk Dataset

"Every cancer screening programme works on one insight: the disease speaks before the patient feels it. CrisisNet applies the same logic to corporate finance."

CrisisNet is a multi-modal, network-aware dataset for building early-warning systems for corporate financial distress. It covers 40 U.S. Energy sector companies (S&P 500) over 10 years (2015–2025) across three parallel signal types: time series financials, NLP text from filings and earnings calls, and a supply-chain network graph.

---

Dataset Structure

The dataset is organised into four top-level folders mirroring the four analytical modules in the CrisisNet architecture:

Module_1/   ← Time Series & Credit Risk (Module A)
Module_2/   ← NLP Text: 10-K Filings + Earnings Calls (Module B)
Module_3/   ← Supply Chain Network Graph (Module C)
Labels/     ← Default & Distress Events (ground truth for all modules)
splits/     ← Pre-computed train / validation / test splits
data/       ← Master company list

---

Module_1 — Time Series & Credit Risk Engine

Purpose: Feeds the X_ts(c,t) feature vector — the financial heartbeat monitor.

Module_1/market_data/

File	Description
all_prices.parquet	Daily OHLCV stock prices for all 40 tickers, 2015–present (2,821 rows × 205 cols)
all_prices.csv	Same as above in CSV format
financials/{TICKER}_income.csv	Quarterly income statement per company
financials/{TICKER}_balance_sheet.csv	Quarterly balance sheet per company
financials/{TICKER}_cashflow.csv	Quarterly cash flow statement per company
financials/{TICKER}_info.csv	Company metadata (sector, market cap, description)

Note: CHK, HES, MRO, PXD, SWN have only _info.csv — these companies were acquired or delisted (Pioneer→Exxon, Hess→Chevron, Marathon Oil→ConocoPhillips, Southwestern→Chesapeake). They are intentionally kept as distress/exit label cases.

Module_1/credit_spreads/

22 FRED series (2005–present) covering credit spreads, treasury yields, macro indicators, and energy prices:

Series	Description
BAMLH0A0HYM2	ICE BofA US High Yield OAS — primary distress signal
BAA10Y, AAA10Y	Moody's corporate bond spreads
VIXCLS	CBOE VIX — market fear gauge
DCOILWTICO, DCOILBRENTEU	WTI & Brent crude oil prices
DHHNGSP	Henry Hub natural gas spot price
T10Y2Y	10Y-2Y Treasury spread (recession predictor)
DGS10, DGS2, DGS3MO	Treasury yields
UNRATE, CPIAUCSL, FEDFUNDS, INDPRO	Macro indicators
fred_all_series.parquet	All 22 series combined (5,681 rows × 22 cols)

Module_1/sec_xbrl/

Structured XBRL financial data from SEC EDGAR for 35/40 companies:

• company_facts/{TICKER}_facts.json — All XBRL-reported line items (Assets, Liabilities, Revenue, EPS, etc.) with full quarterly history
• submissions/{TICKER}_submissions.json — Filing history (dates, form types, accession numbers)
• ticker_cik_mapping.csv — Ticker ↔ SEC CIK number mapping

---

Module_2 — NLP: 10-K Filings & Earnings Calls

Purpose: Feeds the X_nlp(c,t) feature vector via LDA topic modelling and FinBERT sentiment.

Module_2/10k_extracted/10-K/

353 structured JSON files, one per company per year (2015–2024). Each file contains:

{
  "item_1":  "Business description — supply chain, customers, operations...",
  "item_1a": "Risk factors — debt levels, commodity exposure, going concern...",
  "item_7":  "Management Discussion & Analysis — earnings narrative...",
  "item_7a": "Market risk disclosures — interest rate, commodity hedging...",
  "item_8":  "Financial statements narrative..."
}

Naming: {CIK}_{FormType}_{Year}_{AccessionNumber}.json

Usage for NLP:

• Run LDA (Gensim) or BERTopic on item_7 (MD&A) to extract latent distress topics
• Apply FinBERT sentence-by-sentence to item_1a (Risk Factors) for sentiment time series
• Track KL-divergence of topic distributions quarter-over-quarter as a leading signal

Module_2/transcripts/

Earnings call Q&A transcripts from HuggingFace (lamini/earnings-calls-qa):

• 860,164 Q&A records from public company earnings calls
• Fields: question, answer, ticker, date
• Re-download: datasets.load_dataset("lamini/earnings-calls-qa")

---

Module_3 — Supply Chain Network Graph

Purpose: Feeds the X_graph(c,t) feature vector via community detection and contagion simulation.

Module_3/edges_template.csv

30 pre-populated directed edges representing known Energy sector supplier-customer relationships:

source, target, relationship_type, description
SLB, XOM, service_provider, oilfield services
HAL, CVX, service_provider, oilfield services
EPD, VLO, pipeline_supplier, NGL supply
...

Usage: Load into NetworkX → run Louvain community detection → compute DebtRank contagion scores.

Module_3/customer_disclosures_raw.csv

660 customer/supplier disclosure mentions extracted from 10-K Item 1 and Item 7 sections. Use these to augment the graph edges with NLP-extracted relationships.

---

Labels — Default & Distress Events

Purpose: Ground truth labels for all three modules.

Labels/energy_defaults_curated.csv

24 curated bankruptcy/default events (2001–2021):

company, ticker, event_date, event_type, details
Chesapeake Energy, CHK, 2020-06-28, Chapter 11, COVID + legacy debt...
Whiting Petroleum, WLL, 2020-04-01, Chapter 11, COVID oil crash
...

Labels/distress_from_drawdowns.csv

76 mechanically detected distress episodes from stock price drawdowns (>50% peak-to-trough within 6 months) — useful as soft labels for the ML model.

Labels/lopucki_brd_reference.json

Reference pointer to the Florida-UCLA LoPucki Bankruptcy Research Database (1,000+ cases, 1979–2022) for cross-referencing additional default events.

---

Train / Validation / Test Splits

Split strategy: temporal walk-forward (no lookahead leakage)

Split	Period	Rationale
train	2015–2021	Includes 2015–16 oil crash + 2020 COVID wave defaults
validation	2022	Post-COVID recovery, hyperparameter tuning
test	2023–2025	Held-out, never seen during training

Pre-split parquet files are in splits/:

splits/
  stock_prices/     train.parquet, validation.parquet, test.parquet
  fred_macro/       train.parquet, validation.parquet, test.parquet
  labels/
    energy_defaults/    train.parquet, validation.parquet, test.parquet
    distress_drawdowns/ train.parquet, validation.parquet, test.parquet
  10k_filings/      train_manifest.json, validation_manifest.json, test_manifest.json

---

Recommended Usage

Module A — Time Series Credit Risk Engine

import pandas as pd

# Load training data
prices_train = pd.read_parquet("splits/stock_prices/train.parquet")
fred_train   = pd.read_parquet("splits/fred_macro/train.parquet")
labels_train = pd.read_parquet("splits/labels/distress_drawdowns/train.parquet")

# Feature engineering: rolling volatility, Merton Distance-to-Default
# 30-day rolling log-return volatility per ticker
log_ret = prices_train.xs("Close", axis=1, level=0).pct_change().apply(lambda x: (1+x).apply(pd.np.log))
vol_30d = log_ret.rolling(30).std() * (252**0.5)

# Altman Z-Score (benchmark) — requires balance sheet data:
# Z = 1.2*X1 + 1.4*X2 + 3.3*X3 + 0.6*X4 + 1.0*X5
# where X1=Working Capital/TA, X2=Retained Earnings/TA, X3=EBIT/TA,
#       X4=Market Cap/Book Liabilities, X5=Revenue/TA

# Walk-forward cross-validation (never use future data in training window)
# Use expanding window: train on t-36m to t, predict t+1m to t+6m

Module B — NLP Topic Modelling

import json, os
from gensim import corpora, models  # LDA baseline
# from bertopic import BERTopic     # upgraded model

# Load 10-K MD&A sections for training period
train_manifest = json.load(open("splits/10k_filings/train_manifest.json"))
corpus = []
for fpath in train_manifest["files"]:
    filing = json.load(open(fpath))
    text = filing.get("item_7", "") + " " + filing.get("item_1a", "")
    corpus.append(text)

# Preprocessing: tokenise, remove Safe Harbor boilerplate, financial stopwords
# financial_stopwords = ["forward-looking", "may", "could", "believe", "expect", ...]

# Train LDA (K=15 topics typical for earnings text)
# Compare coherence scores across K=10,15,20,25 to find optimal

# FinBERT sentiment on risk factor sentences:
# from transformers import pipeline
# finbert = pipeline("text-classification", model="ProsusAI/finbert")

Module C — Supply Chain Network

import pandas as pd, networkx as nx
import community as community_louvain  # pip install python-louvain

# Build directed graph from known edges
edges = pd.read_csv("Module_3/edges_template.csv")
G = nx.from_pandas_edgelist(edges, "source", "target",
                             edge_attr="relationship_type",
                             create_using=nx.DiGraph())

# Add nodes from company list
companies = pd.read_csv("data/company_list.csv")
for _, row in companies.iterrows():
    G.nodes[row["ticker"]]["subsector"] = row["subsector"]

# Louvain community detection
partition = community_louvain.best_partition(G.to_undirected())

# Centrality metrics
betweenness = nx.betweenness_centrality(G)
pagerank    = nx.pagerank(G)

# DebtRank contagion: mark one node as defaulted, propagate stress
# proportional to edge weights through the graph

Module D — Fusion & Health Score

import lightgbm as lgb
from sklearn.calibration import CalibratedClassifierCV

# Concatenate feature vectors from all three modules:
# X = pd.concat([X_ts, X_nlp, X_graph], axis=1)  # (company × quarter) index

# Walk-forward split (expanding window)
# Train: 2015Q1 → 2020Q4 | Val: 2021Q1 → 2021Q4 | Test: 2022Q1 →

# LightGBM with early stopping on val AUC
model = lgb.LGBMClassifier(n_estimators=500, learning_rate=0.05,
                            num_leaves=31, min_child_samples=10)
model.fit(X_train, y_train,
          eval_set=[(X_val, y_val)],
          eval_metric="auc",
          callbacks=[lgb.early_stopping(50), lgb.log_evaluation(50)])

# Calibrate to produce true probabilities (Platt scaling)
calibrated = CalibratedClassifierCV(model, cv="prefit", method="sigmoid")
calibrated.fit(X_val, y_val)

# SHAP for interpretability
import shap
explainer   = shap.TreeExplainer(model)
shap_values = explainer.shap_values(X_test)
shap.summary_plot(shap_values, X_test)

---

Company Universe

40 S&P 500 Energy sector companies across 6 subsectors:

Subsector	Tickers
Integrated Oil	XOM, CVX, OXY
Exploration & Production	COP, EOG, PXD*, DVN, FANG, MRO*, APA, OVV, HES*, CTRA, MTDR, PR, CHRD
Oilfield Services	SLB, HAL, BKR, FTI, NOV
Refining	VLO, MPC, PSX, DK, PBF
Midstream/Pipelines	KMI, WMB, OKE, ET, EPD, TRGP, DTM, AM
Natural Gas / LNG	EQT, AR, RRC, SWN*, CHK*, LNG

* Delisted/acquired/bankrupt — useful as distress/exit label cases

---

Research Questions (from project proposal)

• RQ1 — Prediction: Can we predict corporate default events 3–6 months in advance using time series + NLP + network features with higher AUC-ROC than Altman Z-Score?
• RQ2 — Contagion: Which companies act as 'super-spreaders' of financial distress — and can community detection identify them before a crisis?
• RQ3 — Narrative Signal: Does sentiment and topic shift in earnings call language provide statistically significant leading signal for credit deterioration?

---

Citation

@dataset{crisisnet2025,
  title     = {CrisisNet: A Multi-Modal Corporate Default Risk Dataset},
  author    = {Sashank and team},
  year      = {2025},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/datasets/Sashank-810/crisisnet-dataset}
}

---

License

CC BY 4.0 — Free to use for research and commercial purposes with attribution.

Data sourced from: Yahoo Finance (yfinance), FRED API (St. Louis Fed), SEC EDGAR (public domain), HuggingFace lamini/earnings-calls-qa.