src/features.py

# src/features.py
# Shared feature engineering — imported by both train.py and predict.py.
# Any change to features must happen here and here only.

import pandas as pd
import numpy as np

# Columns that are identifiers, future leaks, or time proxies.
# train.py and predict.py both reference this list.
DROP_COLS = [
    # identifiers
    "ts", "hhmm", "market_id", "open_mid",
    # future leakage
    "swing_high", "max_mid_reached", "price_range",
    "dist_to_high", "is_last_tick", "outcome", "swing_low",
    # absolute price level — proxies time period, not signal
    "btc_price", "btc_open",
    "prev_open_mid",        # replaced by prev_open_mid_delta below
    # calendar time — regime proxies that don't generalise
    "hour", "minute_of_hour",
    # raw OB levels — replaced by aggregates below
    "bid_s1","bid_s2","bid_s3","bid_s4","bid_s5",
    "ask_s1","ask_s2","ask_s3","ask_s4","ask_s5",
    "bid_p1","bid_p2","bid_p3","bid_p4","bid_p5",
    "ask_p1","ask_p2","ask_p3","ask_p4","ask_p5",
]

# Target and group columns — never used as features
TARGET_COL = "swing_occurred"
GROUP_COL  = "market_seq"


def build_features(df: pd.DataFrame) -> pd.DataFrame:
    """
    Add engineered columns to df in-place (returns a copy).
    Call this on raw CSV data before selecting feature_cols.
    """
    df = df.copy()

    # Aggregate order book sizes → cleaner signal than raw levels
    df["total_bid_size"] = df[["bid_s1","bid_s2","bid_s3","bid_s4","bid_s5"]].sum(axis=1)
    df["total_ask_size"] = df[["ask_s1","ask_s2","ask_s3","ask_s4","ask_s5"]].sum(axis=1)
    df["size_imbalance"] = (
        (df["total_bid_size"] - df["total_ask_size"]) /
        (df["total_bid_size"] + df["total_ask_size"] + 1e-9)
    )

    # OB price slope: how steeply does the book widen on each side?
    df["bid_ask_slope"] = (
        (df["ask_p5"] - df["ask_p1"]) -
        (df["bid_p1"] - df["bid_p5"])
    )

    # Relative shift vs previous market open — avoids absolute price drift
    df["prev_open_mid_delta"] = df["open_mid"] - df["prev_open_mid"]

    return df


def get_entry_snapshot(df: pd.DataFrame) -> pd.DataFrame:
    """
    From a full tick-level DataFrame, return one row per market:
    the first tick where mid enters the entry zone (mid <= 0.35).
    This is the prediction point — what the model sees at trade time.
    """
    df_entry = df[df["mid"] <= 0.35].copy()
    df_model = (
        df_entry
        .sort_values([GROUP_COL, "secs_norm"], ascending=[True, False])
        .groupby(GROUP_COL).first()
        .reset_index()
    )
    return df_model


def get_feature_cols(df: pd.DataFrame) -> list:
    """
    Return the list of feature columns for a given DataFrame,
    after DROP_COLS and target/group columns have been excluded.
    """
    exclude = set(DROP_COLS + [TARGET_COL, GROUP_COL])
    return [c for c in df.columns if c not in exclude]