openavmkit.ratio_study

RatioStudy

RatioStudy(predictions, ground_truth, max_trim)

Performs an IAAO-standard Ratio Study, generating all the relevant statistics.

Attributes:

Name	Type	Description
predictions	ndarray	Series representing predicted values
ground_truth	ndarray	Series representing ground truth values (typically observed sale prices)
count	int	The number of observations
median_ratio	float	The median value of all prediction/ground_truth ratios
mean_ratio	float	The mean value of all prediction/ground_truth ratios
cod	float	The coefficient of dispersion, a measure of variability (lower is better)
cod_trim	float	The coefficient of dispersion, after outlier ratios outside the interquartile range have been trimmed
prd	float	The price-related differential, a measure of vertical equity
prb	float	The price-related bias, a measure of vertical equity

Initialize a ratio study object

Parameters:

Name	Type	Description	Default
predictions	ndarray	Series representing predicted values	required
ground_truth	ndarray	Series representing ground truth values (typically observed sale prices)	required
max_trim	float	The maximum amount of records allowed to be trimmed in a ratio study	required

Source code in openavmkit/ratio_study.py

def __init__(self, predictions: np.ndarray, ground_truth: np.ndarray, max_trim: float):
    """
    Initialize a ratio study object

    Parameters
    ----------
    predictions : np.ndarray
        Series representing predicted values
    ground_truth : np.ndarray
        Series representing ground truth values (typically observed sale prices)
    max_trim : float
        The maximum amount of records allowed to be trimmed in a ratio study
    """
    if len(predictions) != len(ground_truth):
        raise ValueError("predictions and ground_truth must have the same length")

    if len(predictions) == 0:
        self.count = 0
        self.count_trim = 0
        self.predictions = np.array([])
        self.ground_truth = np.array([])
        self.median_ratio = float("nan")
        self.cod = float("nan")
        self.cod_trim = float("nan")
        self.prd = float("nan")
        self.prb = float("nan")
        self.prd_trim = float("nan")
        self.prb_trim = float("nan")
        self.median_ratio_trim = float("nan")
        self.mean_ratio = float("nan")
        self.mean_ratio_trim = float("nan")
        return

    self.count = len(predictions)
    self.predictions = predictions
    self.ground_truth = ground_truth

    ratios = div_series_z_safe(predictions, ground_truth).astype(float)
    if len(ratios) > 0:
        median_ratio = float(np.median(ratios))
    else:
        median_ratio = float("nan")

    # trim the ratios to remove outliers -- trim to the interquartile range
    trim_predictions, trim_ground_truth = stats.trim_outlier_ratios(predictions, ground_truth, max_trim)
    trim_ratios = div_series_z_safe(trim_predictions, trim_ground_truth).astype(float)

    self.count_trim = len(trim_ratios)

    cod = stats.calc_cod(ratios)
    cod_trim = stats.calc_cod(trim_ratios)

    prd = stats.calc_prd(predictions, ground_truth)
    prd_trim = stats.calc_prd(trim_predictions, trim_ground_truth)

    prb, _, _ = stats.calc_prb(predictions, ground_truth)
    prb_trim, _, _ = stats.calc_prb(trim_predictions, trim_ground_truth)

    self.median_ratio = median_ratio

    if len(ratios) == 0:
        self.mean_ratio = float("nan")
    else:
        self.mean_ratio = float(np.mean(ratios))

    if len(trim_ratios) == 0:
        self.mean_ratio_trim = float("nan")
        self.median_ratio_trim = float("nan")
    else:
        self.mean_ratio_trim = float(np.mean(trim_ratios))
        self.median_ratio_trim = float(np.median(trim_ratios))

    self.cod = cod
    self.cod_trim = cod_trim

    self.prd = prd
    self.prd_trim = prd_trim

    self.prb = prb
    self.prb_trim = prb_trim

RatioStudyBootstrapped

RatioStudyBootstrapped(predictions, ground_truth, max_trim, confidence_interval=0.95, iterations=1000)

Performs an IAAO-standard Ratio Study, generating all the relevant statistics. This version adds confidence intervals.

Attributes:

Name	Type	Description
iterations	float	Number of bootstrap iterations
confidence_interval	float	The confidence interval (e.g. 0.95 for 95% confidence)
median_ratio	ConfidenceStat	The median value of all prediction/ground_truth ratios
mean_ratio	ConfidenceStat	The mean value of all prediction/ground_truth ratios
cod	ConfidenceStat	The coefficient of dispersion, a measure of variability (lower is better)
prd	ConfidenceStat	The price-related differential, a measure of vertical equity
median_ratio_trim	ConfidenceStat	The median value of trimmed prediction/ground_truth ratios
mean_ratio_trim	ConfidenceStat	The mean value of trimmed prediction/ground_truth ratios
cod_trim	ConfidenceStat	The coefficient of dispersion, a measure of variability (lower is better), of the trimmed set
prd_trim	ConfidenceStat	The price-related differential, a measure of vertical equity, of the trimmed set

Initialize a Bootstrapped ratio study object

Parameters:

Name	Type	Description	Default
predictions	ndarray	Series representing predicted values	required
ground_truth	ndarray	Series representing ground truth values (typically observed sale prices)	required
max_trim	float	The maximum amount of records allowed to be trimmed in a ratio study	required
confidence_interval	float	Desired confidence interval (default is 0.95, indicating 95% confidence)	0.95
iterations	int	How many bootstrap iterations to perform	1000

Source code in openavmkit/ratio_study.py

def __init__(
    self,
    predictions: np.ndarray,
    ground_truth: np.ndarray,
    max_trim: float,
    confidence_interval: float = 0.95,
    iterations: int = 1000,
):
    """
    Initialize a Bootstrapped ratio study object

    Parameters
    ----------
    predictions : np.ndarray
        Series representing predicted values
    ground_truth : np.ndarray
        Series representing ground truth values (typically observed sale prices)
    max_trim : float
        The maximum amount of records allowed to be trimmed in a ratio study
    confidence_interval : float
        Desired confidence interval (default is 0.95, indicating 95% confidence)
    iterations : int
        How many bootstrap iterations to perform
    """

    self.confidence_interval = confidence_interval

    if len(predictions) == 0:
        self.count = 0
        self.count_trim = 0
        self.iterations = 0
        self.median_ratio = ConfidenceStat(
            value=float("nan"), 
            confidence_interval = confidence_interval,
            low=float("nan"),
            high=float("nan")
        )
        self.mean_ratio = median_ratio = ConfidenceStat(
            value=float("nan"), 
            confidence_interval = confidence_interval,
            low=float("nan"),
            high=float("nan")
        )
        self.cod = ConfidenceStat(
            value=float("nan"), 
            confidence_interval = confidence_interval,
            low=float("nan"),
            high=float("nan")
        )
        self.prd = ConfidenceStat(
            value=float("nan"), 
            confidence_interval = confidence_interval,
            low=float("nan"),
            high=float("nan")
        )
        self.median_ratio_trim = median_ratio = ConfidenceStat(
            value=float("nan"), 
            confidence_interval = confidence_interval,
            low=float("nan"),
            high=float("nan")
        )
        self.mean_ratio_trim = median_ratio = ConfidenceStat(
            value=float("nan"), 
            confidence_interval = confidence_interval,
            low=float("nan"),
            high=float("nan")
        )
        self.cod_trim = ConfidenceStat(
            value=float("nan"), 
            confidence_interval = confidence_interval,
            low=float("nan"),
            high=float("nan")
        )
        self.prd_trim = ConfidenceStat(
            value=float("nan"), 
            confidence_interval = confidence_interval,
            low=float("nan"),
            high=float("nan")
        )
    else:
        self.count = len(ground_truth)
        self.iterations = iterations

        results = stats.calc_ratio_stats_bootstrap(predictions, ground_truth)

        self.cod = results["cod"]
        self.median_ratio = results["median_ratio"]
        self.mean_ratio = results["mean_ratio"]
        self.prd = results["prd"]

        trim_predictions, trim_ground_truth = stats.trim_outlier_ratios(predictions, ground_truth, max_trim)

        self.count_trim = len(trim_ground_truth)

        results = stats.calc_ratio_stats_bootstrap(trim_predictions, trim_ground_truth)

        self.cod_trim = results["cod"]
        self.median_ratio_trim = results["median_ratio"]
        self.mean_ratio_trim = results["mean_ratio"]
        self.prd_trim = results["prd"]

run_and_write_ratio_study_breakdowns

run_and_write_ratio_study_breakdowns(settings)

Runs ratio studies, with breakdowns, and writes them to disk.

Parameters:

Name	Type	Description	Default
settings	dict	Settings dictionary	required

Source code in openavmkit/ratio_study.py

def run_and_write_ratio_study_breakdowns(settings: dict):
    """Runs ratio studies, with breakdowns, and writes them to disk.

    Parameters
    ----------
    settings : dict
        Settings dictionary
    """
    model_groups = get_model_group_ids(settings)
    rs = settings.get("analysis", {}).get("ratio_study", {})
    skip = rs.get("skip", [])
    for model_group in model_groups:
        if model_group in skip:
            print(f"Skipping {model_group}...")
            continue
        print(f"Generating report for {model_group}")
        path = f"out/models/{model_group}/main/model_ensemble.pickle"
        if os.path.exists(path):
            os.makedirs(f"out/models/{model_group}", exist_ok=True)
            ensemble_results = read_pickle(path)
            df_sales = ensemble_results.df_sales
            _run_and_write_ratio_study_breakdowns(
                settings, df_sales, model_group, f"out/models/{model_group}"
            )