Upload scripts/compute_volatility_labels.py with huggingface_hub

scripts/compute_volatility_labels.py

@@ -0,0 +1,178 @@
+# compute_volatility_labels.py
+# Computes lagged_vol_30d and fwd_vol_30d for every filing in filings_index.csv
+# where has_pickle=True.
+#
+# Price sources (in priority order):
+#   1. FINSABER all_sp500_prices_2000_2024_delisted_include.csv  (2000-2024)
+#   2. WRDS CRSP crsp_2025_daily.csv                             (2025)
+#   3. yfinance fallback for ETN / PARA (no PERMNO in CCM)
+#
+# Volatility definition:
+#   vol = std(log_returns over 30 trading days) * sqrt(252)   [annualised]
+#   lagged_vol_30d : 30 trading days strictly BEFORE filing_date
+#   fwd_vol_30d    : 30 trading days strictly AFTER  filing_date
+#
+# Output: datasets/volatility_labels.csv
+# Run:    python scripts/compute_volatility_labels.py
+
+import math
+import numpy as np
+import pandas as pd
+from tqdm import tqdm
+
+# ── Paths ──────────────────────────────────────────────────────────────────────
+FILINGS_PATH  = r"D:\UoE AI\Dissertation\IPP Draft\datasets\filings_index.csv"
+FINSABER_PATH = r"D:\UoE AI\Dissertation\IPP Draft\datasets\all_sp500_prices_2000_2024_delisted_include.csv"
+CRSP_PATH     = r"D:\UoE AI\Dissertation\IPP Draft\datasets\crsp_2025_daily.csv"
+OUT_PATH      = r"D:\UoE AI\Dissertation\IPP Draft\datasets\volatility_labels.csv"
+
+WINDOW    = 30              # trading-day window
+ANNUALIZE = math.sqrt(252)  # annualisation factor
+
+# Tickers with no PERMNO in CCM that have Item 1A pickles -> yfinance fallback
+YFINANCE_TICKERS = {'ETN', 'PARA'}
+
+
+# ── Helper ─────────────────────────────────────────────────────────────────────
+def compute_vol(series: pd.Series, filing_date: pd.Timestamp, window: int = 30):
+    """
+    series : pd.Series, DatetimeIndex ascending, values = daily log-returns
+    Returns (lagged_vol, fwd_vol) as Python floats; np.nan if < window days.
+    """
+    # Ensure 1-D Series (guards against yfinance returning DataFrame column)
+    if isinstance(series, pd.DataFrame):
+        series = series.iloc[:, 0]
+
+    dates  = series.index
+    before = dates[dates < filing_date]
+    after  = dates[dates > filing_date]
+
+    lagged_vol = float(series.loc[before[-window:]].std()) * ANNUALIZE \
+                 if len(before) >= window else np.nan
+    fwd_vol    = float(series.loc[after[:window]].std()) * ANNUALIZE \
+                 if len(after)  >= window else np.nan
+
+    return lagged_vol, fwd_vol
+
+
+# ── 1. Load filings (pickle only) ─────────────────────────────────────────────
+print("Loading filings index...")
+filings = pd.read_csv(FILINGS_PATH)
+filings = filings[filings['has_pickle'] == True].copy()
+filings['filing_date'] = pd.to_datetime(filings['filing_date'])
+print(f"  Filings with pickle : {len(filings):,}")
+
+
+# ── 2. Load FINSABER (2000-2024) ──────────────────────────────────────────────
+print("\nLoading FINSABER prices (2000-2024)...")
+fin = pd.read_csv(FINSABER_PATH, usecols=['date', 'adjusted_close', 'symbol'])
+fin['date']   = pd.to_datetime(fin['date'])
+fin['symbol'] = fin['symbol'].str.upper().str.strip()
+fin = fin.sort_values(['symbol', 'date'])
+
+# Log-return from adjusted close
+fin['log_ret'] = fin.groupby('symbol')['adjusted_close'].transform(
+    lambda x: np.log(x / x.shift(1))
+)
+# Drop NaN and -inf/-inf rows (adjusted_close == 0 produces log(0) = -inf)
+fin = fin[fin['log_ret'].notna() & np.isfinite(fin['log_ret'])][['symbol', 'date', 'log_ret']]
+fin.rename(columns={'symbol': 'ticker'}, inplace=True)
+print(f"  Rows after log-ret  : {len(fin):,}  |  Tickers: {fin['ticker'].nunique()}")
+
+
+# ── 3. Load CRSP 2025 ─────────────────────────────────────────────────────────
+print("\nLoading CRSP 2025 daily...")
+crsp = pd.read_csv(CRSP_PATH, usecols=['Ticker', 'DlyCalDt', 'DlyRet'])
+crsp.columns = ['ticker', 'date', 'dlyret']
+crsp['date']   = pd.to_datetime(crsp['date'])
+crsp['ticker'] = crsp['ticker'].str.upper().str.strip()
+crsp = crsp[crsp['dlyret'].notna()].copy()
+crsp['log_ret'] = np.log(1 + crsp['dlyret'])
+crsp = crsp[['ticker', 'date', 'log_ret']]
+print(f"  Rows               : {len(crsp):,}  |  Tickers: {crsp['ticker'].nunique()}")
+
+
+# ── 4. Combine into a single lookup dict ──────────────────────────────────────
+print("\nMerging price sources...")
+prices = pd.concat([fin, crsp], ignore_index=True)
+prices = (prices
+          .sort_values(['ticker', 'date'])
+          .drop_duplicates(subset=['ticker', 'date']))
+
+# dict: ticker (upper) -> pd.Series(log_ret, index=DatetimeIndex)
+price_dict = {
+    tkr: grp.set_index('date')['log_ret']
+    for tkr, grp in prices.groupby('ticker')
+}
+print(f"  Unique tickers in price dict: {len(price_dict):,}")
+
+
+# ── 5. yfinance fallback for ETN / PARA ───────────────────────────────────────
+yf_needed = YFINANCE_TICKERS & set(filings['ticker'].str.upper())
+if yf_needed:
+    import yfinance as yf
+    print(f"\nFetching yfinance for: {yf_needed}")
+    for tkr in sorted(yf_needed):
+        hist = yf.download(tkr, start='2005-01-01', end='2026-06-01',
+                           auto_adjust=True, progress=False)
+        if hist.empty:
+            print(f"  {tkr}: no data returned")
+            continue
+        # Newer yfinance returns hist['Close'] as a DataFrame; squeeze to Series
+        close = hist['Close']
+        if isinstance(close, pd.DataFrame):
+            close = close.iloc[:, 0]
+        s = np.log(close / close.shift(1)).dropna()
+        s = s[np.isfinite(s)]          # drop any -inf from zero prices
+        s.index = pd.to_datetime(s.index)
+        s.index.name = 'date'
+        s.name = 'log_ret'
+        price_dict[tkr] = s.sort_index()
+        print(f"  {tkr}: {len(s):,} rows  ({s.index[0].date()} -> {s.index[-1].date()})")
+
+
+# ── 6. Compute volatility for every filing ────────────────────────────────────
+print("\nComputing volatility windows...")
+records = []
+
+for _, row in tqdm(filings.iterrows(), total=len(filings), desc="Vol"):
+    ticker       = str(row['ticker']).upper().strip()
+    filing_date  = row['filing_date']
+
+    if ticker in price_dict:
+        lagged_vol, fwd_vol = compute_vol(price_dict[ticker], filing_date, WINDOW)
+    else:
+        lagged_vol, fwd_vol = np.nan, np.nan
+
+    records.append({
+        'ticker':          row['ticker'],
+        'cik':             row['cik'],
+        'sic':             row['sic'],
+        'sic_description': row['sic_description'],
+        'fiscal_year':     row['fiscal_year'],
+        'filing_date':     row['filing_date'].date(),
+        'report_date':     row['report_date'],
+        'lagged_vol_30d':  round(lagged_vol, 6) if pd.notna(lagged_vol) else np.nan,
+        'fwd_vol_30d':     round(fwd_vol,    6) if pd.notna(fwd_vol)    else np.nan,
+    })
+
+df = pd.DataFrame(records)
+df = df.sort_values(['ticker', 'filing_date']).reset_index(drop=True)
+
+
+# ── 7. Summary ────────────────────────────────────────────────────────────────
+total         = len(df)
+has_both      = (df['lagged_vol_30d'].notna() & df['fwd_vol_30d'].notna()).sum()
+lagged_only   = (df['lagged_vol_30d'].notna() & df['fwd_vol_30d'].isna()).sum()
+fwd_only      = (df['lagged_vol_30d'].isna()  & df['fwd_vol_30d'].notna()).sum()
+neither       = (df['lagged_vol_30d'].isna()  & df['fwd_vol_30d'].isna()).sum()
+
+print(f"\nTotal filings processed : {total:,}")
+print(f"  Both vol computed     : {has_both:,}")
+print(f"  Lagged only           : {lagged_only:,}  (fwd window crosses data gap)")
+print(f"  Forward only          : {fwd_only:,}")
+print(f"  Neither (no prices)   : {neither:,}")
+
+df.to_csv(OUT_PATH, index=False)
+print(f"\nSaved -> {OUT_PATH}")
+print(df.head(10).to_string(index=False))