Fix BoxCoxTransform inverse for negative lambda values

The _invert method used `if lmbda > 0` to decide between the power
formula and the exponential. This caused negative lambda values to
incorrectly fall through to the `exp` branch, which is only valid for
lmbda == 0. The Box-Cox inverse for any non-zero lambda is
`(lmbda * y + 1) ** (1 / lmbda)`, regardless of sign.

Changed the condition to `if lmbda != 0` so that negative lambdas
(common for heavy-tailed time series) are inverted correctly.

Author: Mr-Neutr0n

merlion/transform/normalize.py

@@ -200,7 +200,7 @@ def _invert(self, time_series: TimeSeries) -> TimeSeries:
         new_vars = []
         for name, var in time_series.items():
             lmbda = self.lmbda[name]
-            if lmbda > 0:
+            if lmbda != 0:
                 var = (lmbda * var + 1) ** (1 / lmbda)
                 nanvals = var.isna()
                 if nanvals.any():