| """Build the nine model input channels from adjusted 1-minute bars. |
| |
| Reference implementation shipped with the model (pandas and numpy only). The formulas, |
| clips and conventions are documented in preprocessing_contract.md. This file is the |
| executable form of that table. It was checked against the training pipeline's |
| implementation with zero numerical difference. |
| """ |
|
|
| from __future__ import annotations |
|
|
| import numpy as np |
| import pandas as pd |
|
|
| EASTERN = "America/New_York" |
|
|
| CHANNELS = [ |
| "log_return", |
| "hl_range", |
| "oc_body", |
| "log_vol_ratio", |
| "session_flag", |
| "tod_sin", |
| "tod_cos", |
| "dow_sin", |
| "dow_cos", |
| ] |
|
|
|
|
| def make_features(df: pd.DataFrame) -> np.ndarray: |
| """Nine channels for one security. |
| |
| Args: |
| df: one security sorted by `timestamp_eastern`, with columns `open_adj`, |
| `high_adj`, `low_adj`, `close_adj`, `volume_adj`, `timestamp_eastern` |
| (tz-aware; any zone is converted to Eastern, naive is taken as Eastern). |
| |
| Returns: |
| (len(df), 9) float32 array in CHANNELS order. |
| """ |
| ts = df["timestamp_eastern"] |
| if not pd.api.types.is_datetime64_any_dtype(ts): |
| ts = pd.to_datetime(ts) |
| ts = ts.dt.tz_localize(EASTERN) if ts.dt.tz is None else ts.dt.tz_convert(EASTERN) |
| date = ts.dt.date |
| prev_close = df["close_adj"].shift(1) |
| denom = prev_close.fillna(df["close_adj"]) |
|
|
| log_return = np.log(df["close_adj"] / prev_close).fillna(0.0).clip(-0.2, 0.2) |
| hl_range = ((df["high_adj"] - df["low_adj"]) / denom).clip(0.0, 0.2) |
| oc_body = ((df["close_adj"] - df["open_adj"]) / denom).clip(-0.2, 0.2) |
|
|
| daily_mean_vol = df.groupby(date)["volume_adj"].mean() |
| rolling = daily_mean_vol.rolling(20, min_periods=1).mean().shift(1) |
| rolling_per_bar = date.map(rolling) |
| bad_mean = rolling_per_bar.isna() | (rolling_per_bar == 0) |
| ratio = (df["volume_adj"] / rolling_per_bar).where(~bad_mean & (df["volume_adj"] > 0), 1.0) |
| log_vol_ratio = np.clip(np.where(df["volume_adj"] == 0, -10.0, np.log(ratio)), -10.0, 10.0) |
|
|
| session_flag = (date != date.shift(1)).astype(float) |
| tod = (ts.dt.hour + ts.dt.minute / 60.0) * (2 * np.pi / 24.0) |
| dow = ts.dt.weekday * (2 * np.pi / 5.0) |
|
|
| feats = np.stack( |
| [ |
| log_return, |
| hl_range, |
| oc_body, |
| log_vol_ratio, |
| session_flag, |
| np.sin(tod), |
| np.cos(tod), |
| np.sin(dow), |
| np.cos(dow), |
| ], |
| axis=1, |
| ) |
| return feats.astype(np.float32) |
|
|
