docs: update license header in compactts/agent/post_processor.py to BSL-1.1

compactts/agent/post_processor.py

@@ -0,0 +1,64 @@
+"""Calibrated Post-Processor for CompactTS-Agent-v2.
+
+Enforces physical validity, quantile monotonicity, and probabilistic
+interval sharpness for state-of-the-art CRPS calibration.
+
+Attribution: Daksh Bhalala
+License: Business Source License 1.1 (BSL-1.1)
+"""
+
+from typing import Optional
+import numpy as np
+
+
+class CalibratedPostProcessor:
+    """Post-processing pipeline for probabilistic forecasts.
+
+    Ensures:
+    1. Strict quantile non-crossing: q_0.1 <= q_0.2 <= ... <= q_0.9
+    2. Physical support constraints: non-negativity for energy, volume, counts, solar
+    3. Probabilistic quantile calibration via dispersion blending
+    """
+
+    def __init__(self, timestral_quantile_weight: float = 0.0):
+        self.w_tms = timestral_quantile_weight
+
+    def process_quantiles(
+        self,
+        quantiles: np.ndarray,
+        aux_quantiles: Optional[np.ndarray] = None,
+        non_negative: bool = False,
+    ) -> np.ndarray:
+        """Process and calibrate a 9-quantile forecast array.
+
+        Parameters
+        ----------
+        quantiles : np.ndarray
+            Array of shape (9, H) representing quantile levels [0.1..0.9].
+        aux_quantiles : Optional[np.ndarray]
+            Secondary quantile array (e.g., Timestral-209M) for uncertainty calibration.
+        non_negative : bool
+            Whether to clip negative values to 0.0.
+
+        Returns
+        -------
+        np.ndarray
+            Calibrated and monotonic quantile array of shape (9, H).
+        """
+        out = quantiles.copy()
+
+        # If auxiliary quantiles are available, blend for sharp, calibrated uncertainty
+        if aux_quantiles is not None and self.w_tms > 0:
+            median = out[4]  # Preserve point prediction median
+            out = (1.0 - self.w_tms) * out + self.w_tms * aux_quantiles
+            # Shift back so median is anchored
+            out[4] = median
+
+        # Physical non-negativity
+        if non_negative:
+            out = np.maximum(out, 0.0)
+
+        # Enforce strict quantile monotonicity
+        out = np.sort(out, axis=0)
+
+        return out