eth-usd-predictor / src /inference_features.py
Jony Ling
Clean initial commit for HF Space
b717bee
Raw
History Blame Contribute Delete
12.4 kB
"""Live V5 feature builder aligned with the cleaned notebook Part A."""
from pathlib import Path
import numpy as np
import pandas as pd
import polars as pl
EXTERNAL_MARKET_PATHS = [
Path("data/external_market_features.csv"),
Path("data/external_market_features.parquet"),
Path("data/whale_exchange_flow_features.csv"),
]
def _parse_timestamp_column(frame: pl.DataFrame, timestamp_col: str = "timestamp") -> pl.DataFrame:
dtype = frame.schema.get(timestamp_col)
if dtype == pl.Datetime:
return frame.with_columns(pl.col(timestamp_col).dt.cast_time_unit("us").alias(timestamp_col))
return frame.with_columns(
pl.col(timestamp_col).cast(pl.Utf8).str.to_datetime(strict=False).alias(timestamp_col)
).drop_nulls(timestamp_col)
def load_external_market_features(paths=EXTERNAL_MARKET_PATHS, timestamp_col: str = "timestamp") -> pl.DataFrame | None:
frames = []
for raw_path in paths:
path = Path(raw_path)
if not path.exists():
continue
if path.suffix.lower() == ".parquet":
frame = pl.read_parquet(path)
else:
frame = pl.read_csv(path, try_parse_dates=False, infer_schema_length=10000)
if timestamp_col not in frame.columns:
continue
frame = _parse_timestamp_column(frame, timestamp_col).sort(timestamp_col)
numeric_cols = [
col for col, dtype in zip(frame.columns, frame.dtypes)
if col != timestamp_col and dtype.is_numeric()
]
if not numeric_cols:
continue
renamed = {col: f"ext_{col}" if not col.startswith("ext_") else col for col in numeric_cols}
frames.append(frame.select([timestamp_col, *numeric_cols]).rename(renamed))
if not frames:
return None
out = frames[0]
for frame in frames[1:]:
out = out.join_asof(frame, on=timestamp_col, strategy="backward")
return out.sort(timestamp_col)
def add_external_feature_transforms(frame: pl.DataFrame, external_cols: list[str]) -> tuple[pl.DataFrame, list[str]]:
engineered = []
for col in external_cols:
for lag in [1, 4, 8, 24]:
frame = frame.with_columns(pl.col(col).shift(lag).alias(f"{col}_lag_{lag}h"))
engineered.append(f"{col}_lag_{lag}h")
frame = frame.with_columns([
(pl.col(col) - pl.col(col).rolling_mean(24)).alias(f"{col}_dev_24h"),
(pl.col(col) / (pl.col(col).rolling_mean(168) + 1e-10) - 1).alias(f"{col}_vs_168h"),
pl.col(col).diff().rolling_mean(8).alias(f"{col}_flow_8h"),
])
engineered.extend([f"{col}_dev_24h", f"{col}_vs_168h", f"{col}_flow_8h"])
return frame, engineered
def _to_hourly_polars(df_raw: pd.DataFrame) -> pl.DataFrame:
df = df_raw.copy()
if "timestamp" not in df.columns:
if isinstance(df.index, pd.DatetimeIndex):
df = df.reset_index().rename(columns={df.index.name or "index": "timestamp"})
else:
raise ValueError("df_raw must have a timestamp column or DatetimeIndex")
df["timestamp"] = pd.to_datetime(df["timestamp"])
for col, scale in {
"tx_count": 1.5,
"active_senders": 0.8,
"active_receivers": 0.7,
"total_eth_transferred": 1.2,
"total_gas_used": 0.9,
}.items():
if col not in df.columns:
df[col] = df["volume"] * scale
frame = pl.from_pandas(df).with_columns(
pl.col("timestamp").dt.cast_time_unit("us").alias("timestamp")
).sort("timestamp")
diffs = frame["timestamp"].diff().drop_nulls()
is_hourly = len(diffs) > 0 and diffs.dt.total_minutes().median() >= 55
if is_hourly:
return frame.select([
"timestamp", "open", "high", "low", "close", "volume",
"tx_count", "active_senders", "active_receivers",
"total_eth_transferred", "total_gas_used",
])
return frame.group_by_dynamic("timestamp", every="1h").agg([
pl.col("open").first().alias("open"),
pl.col("high").max().alias("high"),
pl.col("low").min().alias("low"),
pl.col("close").last().alias("close"),
pl.col("volume").sum().alias("volume"),
pl.col("tx_count").first().alias("tx_count"),
pl.col("active_senders").first().alias("active_senders"),
pl.col("active_receivers").first().alias("active_receivers"),
pl.col("total_eth_transferred").first().alias("total_eth_transferred"),
pl.col("total_gas_used").first().alias("total_gas_used"),
pl.col("close").count().alias("tick_count_1h"),
]).filter(pl.col("tick_count_1h") >= 30).drop("tick_count_1h").sort("timestamp")
def build_live_v5_features(df_raw: pd.DataFrame) -> tuple[pd.DataFrame, list[str]]:
df_h5 = _to_hourly_polars(df_raw)
external_features = load_external_market_features()
external_feature_cols: list[str] = []
if external_features is not None:
df_h5 = df_h5.join_asof(external_features, on="timestamp", strategy="backward")
external_feature_cols = [col for col in external_features.columns if col != "timestamp"]
df_h5 = df_h5.with_columns([
pl.col(col).fill_null(strategy="forward").fill_null(0.0).alias(col)
for col in external_feature_cols
])
df_h5, external_engineered_cols = add_external_feature_transforms(df_h5, external_feature_cols)
df_h5 = df_h5.with_columns([
pl.col(col).fill_null(strategy="forward").fill_null(0.0).alias(col)
for col in external_engineered_cols
])
external_feature_cols = external_feature_cols + external_engineered_cols
c, h, l, v, o = pl.col("close"), pl.col("high"), pl.col("low"), pl.col("volume"), pl.col("open")
for lag in [1, 2, 4, 6, 12, 24, 48, 168]:
df_h5 = df_h5.with_columns((c / c.shift(lag) - 1).alias(f"return_{lag}h"))
df_h5 = df_h5.with_columns([
((c - o) / (h - l + 1e-10)).alias("candle_body_ratio"),
((h - l) / (c + 1e-10)).alias("range_pct"),
((c - l) / (h - l + 1e-10)).alias("close_location_value"),
((h - c.shift(1)) / (c.shift(1) + 1e-10)).alias("gap_high"),
((l - c.shift(1)) / (c.shift(1) + 1e-10)).alias("gap_low"),
])
for w in [4, 12, 24, 48, 168]:
df_h5 = df_h5.with_columns([
c.rolling_mean(w).alias(f"sma_{w}h"),
c.rolling_std(w).alias(f"vol_{w}h"),
v.rolling_mean(w).alias(f"vol_avg_{w}h"),
(h - l).rolling_mean(w).alias(f"range_avg_{w}h"),
(c / c.rolling_mean(w) - 1).alias(f"price_vs_sma_{w}h"),
])
for period in [6, 14, 24]:
delta = c.diff()
gain = delta.clip(lower_bound=0).rolling_mean(period)
loss = (-delta.clip(upper_bound=0)).rolling_mean(period)
df_h5 = df_h5.with_columns((100 - 100 / (1 + gain / (loss + 1e-10))).alias(f"rsi_{period}h"))
ema12 = c.ewm_mean(span=12)
ema26 = c.ewm_mean(span=26)
macd = ema12 - ema26
macd_signal = macd.ewm_mean(span=9)
df_h5 = df_h5.with_columns([
macd.alias("macd_h"),
macd_signal.alias("macd_signal_h"),
(macd - macd_signal).alias("macd_hist_h"),
])
df_h5 = df_h5.with_columns([
(v / (v.rolling_mean(24) + 1e-10)).alias("vol_ratio_24h"),
(v / (v.rolling_mean(168) + 1e-10)).alias("vol_ratio_168h"),
v.rolling_std(24).alias("vol_volatility_24h"),
(v * (c - c.shift(1)).sign()).rolling_sum(24).alias("obv_24h"),
])
bb_mid = c.rolling_mean(24)
bb_std = c.rolling_std(24)
df_h5 = df_h5.with_columns([
((c - bb_mid) / (bb_std + 1e-10)).alias("bb_zscore_24h"),
(bb_std / (bb_mid + 1e-10)).alias("bb_width_24h"),
])
for col in ["tx_count", "active_senders", "active_receivers", "total_eth_transferred", "total_gas_used"]:
df_h5 = df_h5.with_columns([
(pl.col(col) / (pl.col(col).shift(24) + 1e-10) - 1).alias(f"{col}_change_24h"),
pl.col(col).rolling_mean(24).alias(f"{col}_ma24h"),
pl.col(col).rolling_mean(168).alias(f"{col}_ma168h"),
])
for col in ["tx_count", "active_senders", "total_eth_transferred"]:
df_h5 = df_h5.with_columns(
(pl.col(f"{col}_ma24h") / (pl.col(f"{col}_ma168h") + 1e-10) - 1).alias(f"{col}_momentum")
)
df_h5 = df_h5.with_columns([
(2 * np.pi * pl.col("timestamp").dt.hour() / 24).sin().alias("hour_sin"),
(2 * np.pi * pl.col("timestamp").dt.hour() / 24).cos().alias("hour_cos"),
(2 * np.pi * pl.col("timestamp").dt.weekday() / 7).sin().alias("dow_sin"),
(2 * np.pi * pl.col("timestamp").dt.weekday() / 7).cos().alias("dow_cos"),
])
for lag in [1, 2, 3, 4, 6, 12]:
df_h5 = df_h5.with_columns([
(c / c.shift(1) - 1).shift(lag).alias(f"ret_lag_{lag}"),
pl.col("range_pct").shift(lag).alias(f"range_lag_{lag}"),
])
hr_ret = c / c.shift(1) - 1
vol_short = hr_ret.rolling_std(window_size=24)
vol_long = hr_ret.rolling_std(window_size=720)
df_h5 = df_h5.with_columns([
(vol_short / (vol_long + 1e-10)).alias("vol_regime_ratio"),
vol_long.alias("vol_30d"),
(vol_short - vol_long).alias("vol_shift"),
])
obv = (v * (c - c.shift(1)).sign()).cum_sum()
obv_slope_24 = obv - obv.shift(24)
price_slope_24 = c - c.shift(24)
df_h5 = df_h5.with_columns([
obv_slope_24.alias("obv_slope_24h"),
(obv_slope_24.sign() - price_slope_24.sign()).alias("vol_price_divergence"),
])
avg_xfer = pl.col("total_eth_transferred") / (pl.col("active_senders") + 1e-10)
df_h5 = df_h5.with_columns([
avg_xfer.alias("avg_transfer_size"),
(avg_xfer / (avg_xfer.shift(24) + 1e-10) - 1).alias("transfer_size_change_24h"),
(avg_xfer.rolling_mean(24) / (avg_xfer.rolling_mean(168) + 1e-10) - 1).alias("transfer_size_momentum"),
])
gpv = pl.col("total_gas_used") / (v + 1e-10)
df_h5 = df_h5.with_columns([
gpv.alias("gas_per_volume"),
gpv.rolling_mean(24).alias("gas_per_volume_ma24h"),
(gpv / (gpv.shift(24) + 1e-10) - 1).alias("gas_per_volume_change_24h"),
])
ret_4h = c / c.shift(4) - 1
ret_48h = c / c.shift(48) - 1
ret_168h = c / c.shift(168) - 1
df_h5 = df_h5.with_columns([
(ret_4h - ret_48h).alias("momentum_divergence_4_48"),
(ret_48h - ret_168h).alias("momentum_divergence_48_168"),
(ret_4h.sign() - ret_168h.sign()).alias("trend_alignment"),
vol_short.rolling_std(window_size=48).alias("vol_of_vol_48h"),
])
for period in [6, 14, 24]:
df_h5 = df_h5.with_columns(
(pl.col(f"rsi_{period}h") * pl.col("vol_ratio_24h")).alias(f"rsi{period}_x_vol")
)
df_h5 = df_h5.with_columns([
(pl.col("active_senders") / (pl.col("active_senders").shift(24) + 1e-10) - 1).alias("senders_mom_24h"),
(pl.col("active_senders").rolling_mean(24) / (pl.col("active_senders").rolling_mean(72) + 1e-10) - 1)
.alias("senders_mom_24_vs_72"),
pl.col("active_senders").diff().rolling_mean(24).alias("senders_accel_24h"),
])
df_h5 = df_h5.with_columns([
vol_short.alias("vol_24h_filter"),
(c / c.shift(720) - 1).alias("ret_30d_filter"),
])
vol_arr = df_h5["vol_24h_filter"].to_numpy()
ratio_arr = (df_h5["vol_24h_filter"] / (df_h5["vol_30d"] + 1e-10)).to_numpy()
finite_vol = vol_arr[np.isfinite(vol_arr)]
vol_p25 = np.nanpercentile(finite_vol, 25) if len(finite_vol) else np.nan
vol_p75 = np.nanpercentile(finite_vol, 75) if len(finite_vol) else np.nan
df_h5 = df_h5.with_columns([
pl.Series("vol_regime_low", (vol_arr < vol_p25).astype(np.float64)),
pl.Series("vol_regime_high", (vol_arr > vol_p75).astype(np.float64)),
pl.Series("vol_regime_expanding", (ratio_arr > 1.2).astype(np.float64)),
pl.Series("vol_regime_compressing", (ratio_arr < 0.8).astype(np.float64)),
])
exclude = {
"timestamp", "open", "high", "low", "close", "volume",
"tx_count", "active_senders", "active_receivers",
"total_eth_transferred", "total_gas_used",
"vol_24h_filter", "ret_30d_filter",
}
feature_cols = [col for col in df_h5.columns if col not in exclude and not col.startswith("target_ret_")]
df_feat = df_h5.drop_nulls().to_pandas()
return df_feat, feature_cols