Add beta distribution

[BaerVervergaert]

Added the Beta distribution (scipy.stats.beta) with loc fixed at zero and scale fixed at one.

• The Beta distribution has support for LogScore
• Added Beta LogScore test_beta test

[BaerVervergaert]

Might need to introduce clipping here because sometimes the algorithm overflows and sets value a or b to 0.

[ryan-wolbeck]

@BaerVervergaert can you merge master into your PR when you have time? That way we can test 3.13 as well

[ryan-wolbeck]

    """
    Implements the Beta distribution for NGBoost.
    
    The Beta distribution is defined on the interval [0, 1] and is parameterized
    by two shape parameters a > 0 and b > 0. The distribution is useful for
    modeling bounded continuous data, such as proportions, probabilities, or
    normalized measurements.
    
    Parameters
    ----------
    params : array-like, shape (n_samples, 2)
        Array containing the distribution parameters in log space:
        - params[:, 0]: log(a) - first shape parameter
        - params[:, 1]: log(b) - second shape parameter
    
    Attributes
    ----------
    a : array-like, shape (n_samples,)
        First shape parameter (a > 0), obtained by exponentiating log_a
    b : array-like, shape (n_samples,)
        Second shape parameter (b > 0), obtained by exponentiating log_b
    dist : scipy.stats.beta
        Scipy beta distribution object for sampling and PDF calculations
    log_a : array-like, shape (n_samples,)
        Log of the first shape parameter
    log_b : array-like, shape (n_samples,)
        Log of the second shape parameter

[ryan-wolbeck]

Can we add CRPSScore to be consistent?

[BaerVervergaert]

I'll take a look at it

[ryan-wolbeck]

Something like this might help here

    """
    Initialize Beta distribution with parameters.
    
    Parameters
    ----------
    params : array-like, shape (n_samples, 2)
        Array containing log(a) and log(b) parameters
        
    Raises
    ------
    ValueError
        If params has wrong shape, contains NaN/Inf values, or results in
        non-positive shape parameters
    """
    # Validate input shape
    if len(params) != 2:
        raise ValueError(
            f"Beta distribution requires exactly 2 parameters, got {len(params)}"
        )
    
    # Validate parameter values
    if np.any(np.isnan(params)) or np.any(np.isinf(params)):
        raise ValueError(
            "Invalid parameters: NaN or Inf values detected. "
            "Parameters must be finite numbers."
        )
    
    # Store parameters
    self._params = params
    self.log_a = params[0]
    self.log_b = params[1]
    
    # Convert to shape parameters
    self.a = np.exp(params[0])
    self.b = np.exp(params[1])
    
    # Validate resulting shape parameters
    if np.any(self.a <= 0) or np.any(self.b <= 0):
        raise ValueError(
            "Beta distribution requires positive shape parameters. "
            f"Got a={self.a}, b={self.b}"
        )
    
    # Create scipy distribution object
    self.dist = dist(a=self.a, b=self.b)```

[BaerVervergaert]

Thanks for the feedback :)

It's much appreciated.