fix(ml): add windowed feature flattening for inference parity

The model was trained on 94-candle sliding windows flattened to 2820
features (94 candles x 30 features). Inference was sending raw per-candle
features (27 columns).

Changes:
- Rewrite preprocessing to return (X, window_times) tuple
- Add sliding window creation with correct feature ordering
- Fill missing columns (average, barCount) with 0 for feature parity
- Fill NaN from indicator warmup with 0 instead of dropping rows
- Always compute all indicators (including MFI) for feature parity
- Update predict and batch predict endpoints for new signature
This commit is contained in:
Marko Djordjevic 2026-02-15 22:07:06 +01:00
parent 4c7b3f2676
commit 40d6d1739e
2 changed files with 111 additions and 63 deletions

View file

@ -524,14 +524,8 @@ async def predict(request: PredictRequest):
# Convert candles to list of dicts
candles_data = [candle.model_dump() for candle in request.candles]
# Preprocess candles (feature engineering)
df_preprocessed = preprocess_candles(candles_data, state.pipeline_config)
# Keep times for results mapping
times = df_preprocessed['time'].values
# Extract feature columns (exclude 'time')
X = extract_feature_columns(df_preprocessed)
# Preprocess candles (feature engineering + windowing)
X, window_times = preprocess_candles(candles_data, state.pipeline_config)
# Get predictions and probabilities
if hasattr(state.model, 'predict_proba'):
@ -547,20 +541,18 @@ async def predict(request: PredictRequest):
# Get label names (handle both string and int predictions)
if state.label_encoder is not None:
# Model predicts integers, map to labels
labels = [state.label_encoder.get(int(pred), f"unknown_{pred}") for pred in y_pred]
else:
# Model predicts strings directly
labels = [str(pred) for pred in y_pred]
# Build per-candle predictions
# Build per-window predictions (each window maps to its last candle time)
predictions = [
PredictionResult(
time=int(time),
label=label,
confidence=float(conf)
)
for time, label, conf in zip(times, labels, confidences)
for time, label, conf in zip(window_times, labels, confidences)
]
# Group into spans
@ -577,7 +569,7 @@ async def predict(request: PredictRequest):
)
logger.info(
f"Prediction complete: {len(predictions)} candles, "
f"Prediction complete: {len(predictions)} windows, "
f"{len(spans)} spans, {len([p for p in predictions if p.label != 'O'])} patterns"
)
@ -675,14 +667,8 @@ async def predict_batch(request: BatchPredictRequest):
# Convert batch to candles format
batch_candles = batch_df.to_dict('records')
# Preprocess
df_preprocessed = preprocess_candles(batch_candles, state.pipeline_config)
# Keep times
times = df_preprocessed['time'].values
# Extract features
X = extract_feature_columns(df_preprocessed)
# Preprocess (feature engineering + windowing)
X, window_times = preprocess_candles(batch_candles, state.pipeline_config)
# Predict
if hasattr(state.model, 'predict_proba'):
@ -706,7 +692,7 @@ async def predict_batch(request: BatchPredictRequest):
label=label,
confidence=float(conf)
)
for time, label, conf in zip(times, labels, confidences)
for time, label, conf in zip(window_times, labels, confidences)
]
all_predictions.extend(batch_predictions)

View file

@ -6,7 +6,7 @@ between training and inference.
"""
import logging
from typing import List
from typing import List, Tuple
import pandas as pd
import numpy as np
@ -18,28 +18,40 @@ from features.custom_loader import load_custom_features
logger = logging.getLogger(__name__)
# Window size used during training (number of candles per flattened sample)
TRAINING_WINDOW_SIZE = 94
# Per-candle features expected by the model, in order
TRAINING_FEATURE_ORDER = [
'open', 'high', 'low', 'close', 'volume', 'average', 'barCount',
'rsi_14', 'ema_20', 'ema_50',
'macd_macd', 'macd_signal', 'macd_hist',
'bbands_upper', 'bbands_middle', 'bbands_lower',
'atr_14', 'adx_14', 'cci_14', 'mfi_14',
'stoch_slowk', 'stoch_slowd',
'body_size', 'body_direction', 'upper_wick', 'lower_wick',
'wick_ratio', 'range', 'body_to_range', 'gap',
]
def preprocess_candles(
candles: List[dict],
pipeline_config: PipelineConfig
) -> pd.DataFrame:
) -> Tuple[pd.DataFrame, np.ndarray]:
"""
Preprocess candle data for inference.
Applies the same feature engineering steps as used during training:
1. Convert to DataFrame
2. Validate OHLC data
3. Compute TA-Lib indicators (if enabled)
4. Compute candle features (if enabled)
5. Load custom features (if configured)
6. Drop NaN rows from indicator warmup
Applies the same feature engineering steps as used during training,
then creates sliding windows and flattens them to match training format.
Args:
candles: List of candle dictionaries with time, open, high, low, close, volume
pipeline_config: Pipeline configuration (must match training config)
Returns:
Preprocessed DataFrame ready for model.predict()
Tuple of:
- X: DataFrame with flattened windowed features (one row per window)
- window_times: Array of time values, one per window (time of last candle)
Raises:
ValueError: If data validation fails or too many rows dropped
@ -60,33 +72,28 @@ def preprocess_candles(
except Exception as e:
raise ValueError(f"Candle data validation failed: {e}")
# Handle missing or all-null volume column - fill with 0 if absent/empty
# Handle missing or all-null volume column
if 'volume' not in df.columns or df['volume'].isna().all():
logger.warning("Volume data missing from candles, filling with 0")
df['volume'] = 0.0
# Add missing columns that were present in training data
for col in ['average', 'barCount']:
if col not in df.columns:
df[col] = 0.0
# Get feature engineering config
fe_config = pipeline_config.stages.feature_engineering
if not fe_config.enabled:
logger.warning("Feature engineering disabled in config - returning raw OHLCV")
return df
raise ValueError("Feature engineering must be enabled for windowed inference")
# Compute TA-Lib indicators
# Compute ALL TA-Lib indicators (including volume-dependent ones)
if fe_config.talib_indicators:
indicators = fe_config.talib_indicators
# Skip volume-dependent indicators when volume data is unavailable
volume_indicators = {'MFI', 'OBV', 'AD', 'ADOSC'}
has_real_volume = df['volume'].sum() > 0
if not has_real_volume:
skipped = [i.name for i in indicators if i.name.upper() in volume_indicators]
if skipped:
logger.warning(f"Skipping volume-dependent indicators (no volume data): {skipped}")
indicators = [i for i in indicators if i.name.upper() not in volume_indicators]
logger.info(f"Computing {len(indicators)} TA-Lib indicators")
logger.info(f"Computing {len(fe_config.talib_indicators)} TA-Lib indicators")
try:
df = compute_talib_indicators(df, indicators)
df = compute_talib_indicators(df, fe_config.talib_indicators)
except Exception as e:
logger.error(f"Failed to compute TA-Lib indicators: {e}")
raise ValueError(f"Indicator computation failed: {e}")
@ -109,27 +116,82 @@ def preprocess_candles(
logger.error(f"Failed to load custom features: {e}")
raise ValueError(f"Custom feature loading failed: {e}")
# Handle NaN values from indicator warmup
df_clean = df.dropna()
# Fill NaN values from indicator warmup and missing data with 0
# (instead of dropping rows, since we need contiguous windows)
nan_counts = df.isna().sum()
nan_cols = nan_counts[nan_counts > 0]
if not nan_cols.empty:
logger.info(f"Filling NaN values in {len(nan_cols)} columns (indicator warmup + missing data)")
df = df.fillna(0.0)
rows_dropped = original_rows - len(df_clean)
# Ensure all expected per-candle features exist
for col in TRAINING_FEATURE_ORDER:
if col not in df.columns:
logger.warning(f"Missing expected feature column '{col}', filling with 0")
df[col] = 0.0
if rows_dropped > 0:
logger.info(
f"Dropped {rows_dropped} rows due to indicator warmup "
f"({rows_dropped / original_rows * 100:.1f}%)"
)
logger.info(f"Preprocessing complete: {len(df)} candles with {len(TRAINING_FEATURE_ORDER)} features each")
# Create sliding windows and flatten
X, window_times = create_sliding_windows(df, TRAINING_WINDOW_SIZE, TRAINING_FEATURE_ORDER)
return X, window_times
def create_sliding_windows(
df: pd.DataFrame,
window_size: int,
feature_cols: List[str]
) -> Tuple[np.ndarray, np.ndarray]:
"""
Create sliding windows from per-candle features and flatten.
Each window of `window_size` consecutive candles is flattened into a single
row of features: [feat0_candle0, feat1_candle0, ..., featN_candle0,
feat0_candle1, ..., featN_candleM]
Args:
df: DataFrame with per-candle features and 'time' column
window_size: Number of candles per window
feature_cols: Ordered list of feature column names
# Warn if too much data was lost
if rows_dropped / original_rows > 0.5:
raise ValueError(
f"More than 50% of candles dropped due to indicator warmup "
f"({rows_dropped}/{original_rows}). Provide more historical candles."
)
Returns:
Tuple of:
- X: numpy array of shape (n_windows, window_size * n_features)
- window_times: array of time values (last candle time in each window)
"""
n_candles = len(df)
n_features = len(feature_cols)
logger.info(f"Preprocessing complete: {len(df_clean)} candles ready for prediction")
if n_candles < window_size:
raise ValueError(
f"Not enough candles ({n_candles}) for window size {window_size}. "
f"Need at least {window_size} candles."
)
return df_clean
# Extract feature matrix in correct column order
feature_matrix = df[feature_cols].values # shape: (n_candles, n_features)
times = df['time'].values
n_windows = n_candles - window_size + 1
# Create flattened windows using stride tricks for efficiency
# Each window: candle features are interleaved as col_0, col_1, ..., col_N for each candle index
X = np.zeros((n_windows, window_size * n_features), dtype=np.float64)
window_times = np.zeros(n_windows, dtype=times.dtype)
for i in range(n_windows):
window = feature_matrix[i:i + window_size] # shape: (window_size, n_features)
# Flatten: row-major means [candle0_feat0, candle0_feat1, ..., candle1_feat0, ...]
# But training used {col}_{candle_idx} ordering, which is column-first per candle
# i.e., open_0, high_0, ..., gap_0, open_1, high_1, ..., gap_1, ...
X[i] = window.flatten() # row-major: candle0_all_feats, candle1_all_feats, ...
window_times[i] = times[i + window_size - 1] # last candle in window
logger.info(f"Created {n_windows} sliding windows of size {window_size} "
f"({n_windows * n_features * window_size} total features)")
return X, window_times
def extract_feature_columns(