| """Advanced Feature Engineering Part 1 - Microstructure & Cross-Sectional""" |
| import numpy as np |
| import pandas as pd |
| from typing import Dict, List, Optional |
| import warnings |
| warnings.filterwarnings('ignore') |
|
|
|
|
| class MicrostructureFeatures: |
| """Market microstructure features from OHLCV data""" |
| |
| @staticmethod |
| def amihud_illiquidity(close, volume, window=21): |
| """Amihud (2002) illiquidity: avg |return| / dollar_volume""" |
| dollar_vol = close * volume |
| abs_ret = close.pct_change().abs() |
| return (abs_ret / dollar_vol).rolling(window).mean() |
| |
| @staticmethod |
| def kyle_lambda(close, volume, window=21): |
| """Kyle's lambda: price impact per unit volume""" |
| abs_ret = close.pct_change().abs() |
| signed_vol = volume * np.sign(close.pct_change()) |
| cov = abs_ret.rolling(window).cov(signed_vol) |
| var = signed_vol.rolling(window).var() |
| return cov / var.replace(0, np.nan) |
| |
| @staticmethod |
| def bid_ask_spread_proxy(high, low, close, window=21): |
| """Corwin & Schultz (2012) spread estimator""" |
| beta = ((high - low) ** 2).rolling(window).sum() |
| spread = 2 * (np.exp(np.sqrt(2*beta) - beta) - 1) / (1 + np.exp(np.sqrt(2*beta) - beta)) |
| return spread |
| |
| @staticmethod |
| def vwap(close, high, low, volume, window=14): |
| """Volume-weighted average price""" |
| typical_price = (high + low + close) / 3 |
| vp = typical_price * volume |
| cum_vp = vp.rolling(window).sum() |
| cum_vol = volume.rolling(window).sum() |
| return cum_vp / cum_vol.replace(0, np.nan) |
| |
| @staticmethod |
| def roll_spread(close, window=20): |
| """Roll (1984) effective spread estimator""" |
| delta_p = close.diff() |
| cov = delta_p.rolling(window).apply(lambda x: np.cov(x[:-1], x[1:])[0,1]) |
| return 2 * np.sqrt(-cov.clip(upper=0)) |
| |
| @staticmethod |
| def compute_all(close, high, low, volume): |
| """Compute all microstructure features""" |
| features = pd.DataFrame(index=close.index) |
| features['amihud_illiquidity'] = MicrostructureFeatures.amihud_illiquidity(close, volume) |
| features['kyle_lambda'] = MicrostructureFeatures.kyle_lambda(close, volume) |
| features['bid_ask_spread'] = MicrostructureFeatures.bid_ask_spread_proxy(high, low, close) |
| features['vwap_ratio'] = close / MicrostructureFeatures.vwap(close, high, low, volume) |
| features['roll_spread'] = MicrostructureFeatures.roll_spread(close) |
| |
| dollar_vol = close * volume |
| features['dollar_vol_rank'] = dollar_vol.rolling(63).rank(pct=True) |
| features['volume_trend'] = volume.rolling(21).mean() / volume.rolling(63).mean() - 1 |
| features['volume_delta'] = ((close > close.shift(1)).astype(float) * volume - \ |
| (close < close.shift(1)).astype(float) * volume) / volume.rolling(21).mean() |
| return features |
|
|
|
|
| class CrossSectionalFeatures: |
| """Cross-sectional ranking and momentum features""" |
| |
| @staticmethod |
| def momentum_score(returns, periods=[5, 21, 63, 126, 252]): |
| """Cross-sectional momentum ranking""" |
| features = pd.DataFrame(index=returns.index) |
| for p in periods: |
| cum_ret = returns.rolling(p).sum() |
| features[f'cs_mom_{p}d'] = cum_ret.rank(axis=1, pct=True) |
| return features |
| |
| @staticmethod |
| def mean_reversion(returns, short=5, long=63): |
| """Short-term reversal vs medium-term momentum""" |
| short_ret = returns.rolling(short).sum() |
| long_ret = returns.rolling(long).sum() |
| features = pd.DataFrame(index=returns.index) |
| features['mr_signal'] = short_ret.rank(axis=1, pct=True) - long_ret.rank(axis=1, pct=True) |
| features['mr_short'] = -short_ret.rank(axis=1, pct=True) |
| return features |
| |
| @staticmethod |
| def dispersion(returns, window=21): |
| """Cross-sectional return dispersion""" |
| features = pd.DataFrame(index=returns.index) |
| features['cs_std'] = returns.rolling(window).std(axis=1) |
| features['cs_range'] = returns.rolling(window).max(axis=1) - returns.rolling(window).min(axis=1) |
| features['cs_skew'] = returns.rolling(window).skew(axis=1) |
| features['cs_kurt'] = returns.rolling(window).kurt(axis=1) |
| return features |
|
|
