P2-ETF-TFT-PREDICTOR / data_manager.py
P2SAMAPA
Update data_manager.py
b49a36f unverified
"""
Data fetching, processing, and feature engineering
"""
import pandas as pd
import numpy as np
import yfinance as yf
import streamlit as st
from datetime import datetime
from huggingface_hub import HfApi
import os
import requests
import time
from utils import get_est_time
REPO_ID = "P2SAMAPA/my-etf-data"
# ── ETF universe ──────────────────────────────────────────────────────────────
# TBT removed (leveraged decay). Added: VCIT, LQD, HYG (investment grade + HY credit)
ETF_LIST = ["TLT", "VCIT", "LQD", "HYG", "VNQ", "SLV", "GLD", "AGG", "SPY"]
TARGET_ETFS = ["TLT", "VCIT", "LQD", "HYG", "VNQ", "SLV", "GLD"] # excludes benchmarks
# ── FRED fetch (no pandas_datareader dependency) ──────────────────────────────
def _fetch_fred_series(series_id: str, start_date: str = "2008-01-01") -> pd.Series:
"""
Fetch a single FRED series directly via the public CSV endpoint.
No API key or pandas_datareader required.
"""
try:
url = f"https://fred.stlouisfed.org/graph/fredgraph.csv?id={series_id}"
df = pd.read_csv(url, index_col=0, parse_dates=True)
df.index.name = "Date"
s = df.iloc[:, 0].replace(".", float("nan"))
s = pd.to_numeric(s, errors="coerce").dropna()
s = s[s.index >= pd.Timestamp(start_date)]
s.name = series_id
return s
except Exception as e:
log_warn(f"FRED fetch failed for {series_id}: {e}")
return pd.Series(name=series_id, dtype=float)
def _fetch_fred_multi(series_map: dict, start_date: str = "2008-01-01") -> pd.DataFrame:
"""
Fetch multiple FRED series and return as aligned DataFrame.
series_map: {fred_id: column_name}
"""
frames = []
for series_id, col_name in series_map.items():
s = _fetch_fred_series(series_id, start_date=start_date)
if not s.empty:
s.name = col_name
frames.append(s)
time.sleep(0.2) # polite pacing
if not frames:
return pd.DataFrame()
df = pd.concat(frames, axis=1)
if df.index.tz is not None:
df.index = df.index.tz_localize(None)
return df.sort_index().ffill(limit=5)
def log_warn(msg):
"""Warning that works both in Streamlit and headless mode."""
try:
st.warning(msg)
except Exception:
import logging
logging.getLogger(__name__).warning(msg)
def log_error(msg):
try:
st.error(msg)
except Exception:
import logging
logging.getLogger(__name__).error(msg)
def fetch_macro_data_robust(start_date="2008-01-01"):
"""Fetch macro signals from FRED (direct CSV) and Yahoo Finance"""
all_data = []
# 1. FRED β€” direct CSV fetch, no pandas_datareader needed
try:
fred_symbols = {
"T10Y2Y": "T10Y2Y",
"T10Y3M": "T10Y3M",
"DTB3": "DTB3",
"BAMLH0A0HYM2": "HY_Spread",
"VIXCLS": "VIX",
"DTWEXBGS": "DXY",
}
fred_data = _fetch_fred_multi(fred_symbols, start_date=start_date)
if not fred_data.empty:
all_data.append(fred_data)
else:
log_warn("⚠️ FRED returned empty data")
except Exception as e:
log_warn(f"⚠️ FRED fetch failed: {e}")
# 2. Yahoo Finance β€” gold, copper, VIX
try:
yf_symbols = {"GC=F": "GOLD", "HG=F": "COPPER", "^VIX": "VIX_YF"}
yf_data = yf.download(
list(yf_symbols.keys()), start=start_date, progress=False, auto_adjust=True
)["Close"]
if isinstance(yf_data, pd.Series):
yf_data = yf_data.to_frame()
yf_data.columns = [yf_symbols.get(str(col), str(col)) for col in yf_data.columns]
if yf_data.index.tz is not None:
yf_data.index = yf_data.index.tz_localize(None)
all_data.append(yf_data)
except Exception as e:
log_warn(f"⚠️ Yahoo Finance failed: {e}")
# 3. VIX term structure
try:
vix_term = yf.download(
["^VIX", "^VIX3M"], start=start_date, progress=False, auto_adjust=True
)["Close"]
if not vix_term.empty:
if isinstance(vix_term, pd.Series):
vix_term = vix_term.to_frame()
vix_term.columns = ["VIX_Spot", "VIX_3M"]
vix_term["VIX_Term_Slope"] = vix_term["VIX_3M"] - vix_term["VIX_Spot"]
if vix_term.index.tz is not None:
vix_term.index = vix_term.index.tz_localize(None)
all_data.append(vix_term)
except Exception as e:
log_warn(f"⚠️ VIX Term Structure failed: {e}")
if all_data:
combined = pd.concat(all_data, axis=1, join="outer")
combined = combined.loc[:, ~combined.columns.duplicated()]
combined = combined.ffill(limit=5)
return combined
else:
log_error("❌ Failed to fetch any macro data!")
return pd.DataFrame()
def fetch_etf_data(etfs, start_date="2008-01-01"):
"""
Fetch ETF price data and calculate features.
Produces exactly 3 columns per ETF to match HF dataset schema:
{ETF}_Ret β€” daily return
{ETF}_MA20 β€” 20-day simple moving average of price (raw)
{ETF}_Vol β€” 20-day annualised realised volatility
"""
try:
etf_data = yf.download(
etfs, start=start_date, progress=False, auto_adjust=True
)["Close"]
if isinstance(etf_data, pd.Series):
etf_data = etf_data.to_frame()
if etf_data.index.tz is not None:
etf_data.index = etf_data.index.tz_localize(None)
daily_rets = etf_data.pct_change()
# Daily returns
etf_returns = daily_rets.copy()
etf_returns.columns = [f"{col}_Ret" for col in etf_returns.columns]
# 20-day simple moving average (raw price)
etf_ma20 = etf_data.rolling(20).mean()
etf_ma20.columns = [f"{col}_MA20" for col in etf_ma20.columns]
# 20-day annualised realised volatility
etf_vol = daily_rets.rolling(20).std() * np.sqrt(252)
etf_vol.columns = [f"{col}_Vol" for col in etf_vol.columns]
result = pd.concat([etf_returns, etf_ma20, etf_vol], axis=1)
return result
except Exception as e:
log_error(f"❌ ETF fetch failed: {e}")
return pd.DataFrame()
def smart_update_hf_dataset(new_data, token, force_upload=False):
"""
Smart update: merges new_data on top of existing HF dataset and uploads
if anything changed β€” or always uploads when force_upload=True.
Also handles newly added ETFs: detects ETFs whose _Ret column is missing
or all-NaN in the existing dataset, fetches their full history back to
2008, and backfills before uploading.
"""
if not token:
log_warn("⚠️ No HF_TOKEN found. Skipping dataset update.")
return new_data
raw_url = f"https://huggingface.co/datasets/{REPO_ID}/resolve/main/etf_data.csv"
try:
# Cache-bust the HF CDN so we always read the latest version
bust_url = f"{raw_url}?t={int(time.time())}"
existing_df = pd.read_csv(bust_url)
existing_df.columns = existing_df.columns.str.strip()
date_col = next(
(c for c in existing_df.columns if c.lower() in ["date", "unnamed: 0"]),
existing_df.columns[0],
)
existing_df[date_col] = pd.to_datetime(existing_df[date_col])
existing_df = existing_df.set_index(date_col).sort_index()
if existing_df.index.tz is not None:
existing_df.index = existing_df.index.tz_localize(None)
# ── Step 1: fetch FULL ETF history from 2008 ───────────────────────
full_etf = fetch_etf_data(ETF_LIST, start_date="2008-01-01")
if full_etf.index.tz is not None:
full_etf.index = full_etf.index.tz_localize(None)
# Detect new ETFs for reporting
new_etf_cols = [
etf for etf in ETF_LIST
if f"{etf}_Ret" not in existing_df.columns
or existing_df[f"{etf}_Ret"].isna().mean() > 0.9
]
# ── Step 2: extract macro columns from existing_df ───────────────────
etf_col_names = [c for c in existing_df.columns
if any(c.startswith(f"{e}_") for e in
["TLT","TBT","VCIT","LQD","HYG","VNQ","SLV","GLD","AGG","SPY"])]
macro_col_names = [c for c in existing_df.columns if c not in etf_col_names]
macro_existing = existing_df[macro_col_names].copy()
# ── Step 3: build combined from scratch using pd.concat ──────────────
macro_new_cols = [c for c in new_data.columns
if not any(c.startswith(f"{e}_") for e in
["TLT","TBT","VCIT","LQD","HYG","VNQ","SLV","GLD","AGG","SPY"])]
macro_new = new_data[macro_new_cols].copy() if macro_new_cols else pd.DataFrame()
if not macro_new.empty:
macro_combined = macro_new.combine_first(macro_existing)
else:
macro_combined = macro_existing
# Align on union of all dates
full_index = full_etf.index.union(macro_combined.index)
etf_aligned = full_etf.reindex(full_index)
macro_aligned = macro_combined.reindex(full_index)
combined = pd.concat([etf_aligned, macro_aligned], axis=1)
# ── Step 4: decide whether to upload ─────────────────────────────────
new_rows = len(combined) - len(existing_df)
old_nulls = existing_df.isna().sum().sum()
new_nulls = combined.isna().sum().sum()
filled_gaps = old_nulls - new_nulls
needs_update = force_upload or new_rows > 0 or filled_gaps > 0 or len(new_etf_cols) > 0
if needs_update:
combined.index.name = "Date"
missing = [e for e in ["VCIT","LQD","HYG"] if f"{e}_Ret" not in combined.columns]
present = [e for e in ["VCIT","LQD","HYG"] if f"{e}_Ret" in combined.columns]
try:
st.info(f"πŸ“‹ Pre-upload check β€” new ETFs present: {present} | missing: {missing}")
if present:
sample = combined[[f"{e}_Ret" for e in present]].dropna().head(3)
st.info(f"πŸ“‹ Sample data:\n{sample.to_string()}")
except Exception:
pass
out_df = combined.reset_index()
csv_filename = f"etf_data_{int(time.time())}.csv"
out_df.to_csv(csv_filename, index=False)
api = HfApi()
commit_msg = (
("FORCE " if force_upload else "") +
f"Update: {get_est_time().strftime('%Y-%m-%d %H:%M EST')} | "
f"+{new_rows} rows, filled {filled_gaps} gaps" +
(f", backfilled {new_etf_cols}" if new_etf_cols else "")
)
from huggingface_hub import CommitOperationAdd
with open(csv_filename, "rb") as f:
file_bytes = f.read()
try:
st.info(f"πŸ“€ Uploading {len(file_bytes):,} bytes to HF...")
except Exception:
pass
operations = [CommitOperationAdd(
path_in_repo="etf_data.csv",
path_or_fileobj=file_bytes,
)]
api.create_commit(
repo_id=REPO_ID,
repo_type="dataset",
token=token,
commit_message=commit_msg,
operations=operations,
)
try:
st.success(f"βœ… Dataset updated: +{new_rows} rows, filled {filled_gaps} gaps"
+ (f", backfilled {new_etf_cols}" if new_etf_cols else ""))
except Exception:
pass
return combined
else:
try:
st.info("πŸ“Š Dataset already up-to-date. No upload needed.")
except Exception:
pass
return existing_df
except Exception as e:
log_warn(f"⚠️ Dataset update failed: {e}. Using new data only.")
return new_data
def add_regime_features(df):
"""Add regime detection features using pd.concat to avoid fragmentation"""
new_cols = {}
if "VIX" in df.columns:
new_cols["VIX_Regime_Low"] = (df["VIX"] < 15).astype(int)
new_cols["VIX_Regime_Med"] = ((df["VIX"] >= 15) & (df["VIX"] < 25)).astype(int)
new_cols["VIX_Regime_High"] = (df["VIX"] >= 25).astype(int)
if "T10Y2Y" in df.columns:
new_cols["YC_Inverted"] = (df["T10Y2Y"] < 0).astype(int)
new_cols["YC_Flat"] = ((df["T10Y2Y"] >= 0) & (df["T10Y2Y"] < 0.5)).astype(int)
new_cols["YC_Steep"] = (df["T10Y2Y"] >= 0.5).astype(int)
if "HY_Spread" in df.columns:
new_cols["Credit_Stress_Low"] = (df["HY_Spread"] < 400).astype(int)
new_cols["Credit_Stress_Med"] = ((df["HY_Spread"] >= 400) & (df["HY_Spread"] < 600)).astype(int)
new_cols["Credit_Stress_High"] = (df["HY_Spread"] >= 600).astype(int)
if "VIX_Term_Slope" in df.columns:
new_cols["VIX_Term_Contango"] = (df["VIX_Term_Slope"] > 2).astype(int)
new_cols["VIX_Term_Backwardation"] = (df["VIX_Term_Slope"] < -2).astype(int)
if "T10Y3M" in df.columns:
new_cols["Rates_VeryLow"] = (df["T10Y3M"] < 1.0).astype(int)
new_cols["Rates_Low"] = ((df["T10Y3M"] >= 1.0) & (df["T10Y3M"] < 2.0)).astype(int)
new_cols["Rates_Normal"] = ((df["T10Y3M"] >= 2.0) & (df["T10Y3M"] < 3.0)).astype(int)
new_cols["Rates_High"] = (df["T10Y3M"] >= 3.0).astype(int)
if "T10Y2Y" in df.columns:
yc_mom20 = df["T10Y2Y"].diff(20)
yc_mom60 = df["T10Y2Y"].diff(60)
new_cols["YC_Mom20d"] = yc_mom20
new_cols["YC_Mom60d"] = yc_mom60
new_cols["Rates_Rising20d"] = (yc_mom20 > 0).astype(int)
new_cols["Rates_Falling20d"] = (yc_mom20 < 0).astype(int)
new_cols["Rates_Rising60d"] = (yc_mom60 > 0).astype(int)
new_cols["Rates_Falling60d"] = (yc_mom60 < 0).astype(int)
yc_accel = yc_mom20.diff(20)
new_cols["YC_Accel"] = yc_accel
new_cols["Rates_Accelerating"] = (yc_accel > 0).astype(int)
if "T10Y3M" in df.columns:
t3m_mom20 = df["T10Y3M"].diff(20)
t3m_mom60 = df["T10Y3M"].diff(60)
new_cols["T10Y3M_Mom20d"] = t3m_mom20
new_cols["T10Y3M_Mom60d"] = t3m_mom60
new_cols["T10Y3M_Rising20d"] = (t3m_mom20 > 0).astype(int)
new_cols["T10Y3M_Falling20d"] = (t3m_mom20 < 0).astype(int)
new_cols["T10Y3M_Rising60d"] = (t3m_mom60 > 0).astype(int)
new_cols["T10Y3M_Falling60d"] = (t3m_mom60 < 0).astype(int)
if new_cols:
df = pd.concat([df, pd.DataFrame(new_cols, index=df.index)], axis=1)
return df
def get_data(start_year, force_refresh=False, clean_hf_dataset=False):
"""Main data fetching and processing pipeline"""
raw_url = f"https://huggingface.co/datasets/{REPO_ID}/resolve/main/etf_data.csv"
df = pd.DataFrame()
# ── Load from HuggingFace ─────────────────────────────────────────────────
try:
df = pd.read_csv(f"{raw_url}?t={int(time.time())}")
df.columns = df.columns.str.strip()
date_col = next(
(c for c in df.columns if c.lower() in ["date", "unnamed: 0"]), df.columns[0]
)
df[date_col] = pd.to_datetime(df[date_col])
df = df.set_index(date_col).sort_index()
if df.index.tz is not None:
df.index = df.index.tz_localize(None)
# Optional: clean >30% NaN columns
if clean_hf_dataset:
try:
st.warning("🧹 **Cleaning HF Dataset Mode Active**")
except Exception:
pass
original_cols = len(df.columns)
nan_pct = (df.isna().sum() / len(df)) * 100
bad_cols = nan_pct[nan_pct > 30].index.tolist()
if bad_cols:
df = df.drop(columns=bad_cols)
token = os.getenv("HF_TOKEN")
if token:
df.index.name = "Date"
csv_filename = f"etf_data_{int(time.time())}.csv"
df.reset_index().to_csv(csv_filename, index=False)
api = HfApi()
from huggingface_hub import CommitOperationAdd
with open(csv_filename, "rb") as f:
file_bytes = f.read()
api.create_commit(
repo_id=REPO_ID,
repo_type="dataset",
token=token,
commit_message=f"Cleaned dataset: Removed {len(bad_cols)} columns "
f"({original_cols} β†’ {len(df.columns)})",
operations=[CommitOperationAdd(
path_in_repo="etf_data.csv",
path_or_fileobj=file_bytes,
)],
)
try:
st.success("βœ… HF dataset updated!")
except Exception:
pass
except Exception as e:
log_warn(f"⚠️ Could not load from HuggingFace: {e}")
# ── Sync / force refresh ──────────────────────────────────────────────────
from utils import is_sync_window
should_sync = is_sync_window() or force_refresh
if should_sync:
sync_reason = "πŸ”„ Manual Refresh" if force_refresh else "πŸ”„ Sync Window Active"
try:
ctx = st.status(f"{sync_reason} - Updating Dataset...", expanded=False)
except Exception:
from contextlib import nullcontext
ctx = nullcontext()
with ctx:
etf_data = fetch_etf_data(ETF_LIST)
macro_data = fetch_macro_data_robust()
if not etf_data.empty and not macro_data.empty:
new_df = pd.concat([etf_data, macro_data], axis=1)
token = os.getenv("HF_TOKEN")
df = smart_update_hf_dataset(new_df, token, force_upload=force_refresh)
# ── Fallback: fetch fresh if still empty ──────────────────────────────────
if df.empty:
log_warn("πŸ“Š Fetching fresh data...")
etf_data = fetch_etf_data(ETF_LIST)
macro_data = fetch_macro_data_robust()
if not etf_data.empty and not macro_data.empty:
df = pd.concat([etf_data, macro_data], axis=1)
# ── Feature engineering: Z-scores ────────────────────────────────────────
macro_cols = [
"VIX", "DXY", "COPPER", "GOLD", "HY_Spread", "T10Y2Y", "T10Y3M",
"VIX_Spot", "VIX_3M", "VIX_Term_Slope",
]
for col in df.columns:
if any(m in col for m in macro_cols) or "_Vol" in col:
roll_mean = df[col].rolling(20, min_periods=5).mean()
roll_std = df[col].rolling(20, min_periods=5).std()
df[f"{col}_Z"] = (df[col] - roll_mean) / (roll_std + 1e-9)
# ── Regime features ───────────────────────────────────────────────────────
try:
st.write("🎯 **Adding Regime Detection Features...**")
except Exception:
pass
df = add_regime_features(df)
# ── Filter by start year ──────────────────────────────────────────────────
df = df[df.index.year >= start_year]
try:
st.info(f"πŸ“… After year filter ({start_year}+): {len(df)} samples")
except Exception:
pass
# ── Drop columns with >50% NaNs ───────────────────────────────────────────
nan_pct = df.isna().sum() / len(df)
bad_features = nan_pct[nan_pct > 0.5].index.tolist()
if bad_features:
df = df.drop(columns=bad_features)
# ── Fill remaining NaNs ───────────────────────────────────────────────────
df = df.ffill(limit=5).bfill(limit=100).ffill()
df = df.dropna()
if len(df) > 0:
try:
st.success(
f"βœ… Final dataset: {len(df)} samples from "
f"{df.index[0].strftime('%Y-%m-%d')} to {df.index[-1].strftime('%Y-%m-%d')}"
)
except Exception:
pass
return df