Uncertainty Visualizations

This page showcases examples of plots specifically designed for exploring, diagnosing, and communicating aspects of predictive uncertainty using k-diagram.

Note

You need to run the code snippets locally to generate the plot images referenced below (e.g., ../images/gallery_actual_vs_predicted.png). Ensure the image paths in the .. image:: directives match where you save the plots (likely an images subdirectory relative to this file, e.g., ../images/).

Actual vs. Predicted

Compares actual observed values against point predictions (e.g., Q50) sample-by-sample. Useful for assessing basic accuracy and bias.

import kdiagram as kd
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt

# --- Data Generation ---
np.random.seed(66)
n_points = 120
df = pd.DataFrame({'sample': range(n_points)})
signal = 20 + 15 * np.cos(np.linspace(0, 6 * np.pi, n_points))
df['actual'] = signal + np.random.randn(n_points) * 3
df['predicted'] = signal * 0.9 + np.random.randn(n_points) * 2 + 2

# --- Plotting ---
kd.plot_actual_vs_predicted(
    df=df,
    actual_col='actual',
    pred_col='predicted',
    title='Gallery: Actual vs. Predicted (Dots)',
    line=False, # Use dots instead of lines
    r_label="Value",
    actual_props={'s': 25, 'alpha': 0.7, 'color':'black'}, # Explicit color
    pred_props={'s': 35, 'marker': 'x', 'alpha': 0.7, 'color':'red'}, # Explicit color & size
    # Save the plot (adjust path relative to docs/source/)
    savefig="gallery/images/gallery_actual_vs_predicted.png"
)
plt.close() # Close the plot window after saving

🧠 Analysis and Interpretation

This polar plot provides a direct visual comparison between actual values (black dots) and model-predicted medians (red crosses, Q50) across a set of samples arranged angularly (by index). Each point’s distance from the center (radius) corresponds to the magnitude of its value.

Key Insights:

• Accuracy & Discrepancies: Close alignment between black dots and red crosses indicates accurate predictions for that sample. Deviations highlight errors. The grey connecting lines (if line=True) emphasize the error magnitude and direction.

• Systematic Bias: Look for consistent patterns where red crosses are generally inside (under-prediction) or outside (over-prediction) the black dots.

• Outliers: Samples with unusually large gaps between actual and predicted values are easily spotted.

🔍 In this Example:

• The points form clear cyclical patterns, matching the underlying cosine wave used in data generation.

• Predictions (red crosses) generally track the actual values (black dots) but exhibit some scatter (noise) and slight magnitude differences, particularly near the peaks (outer radius) and troughs (inner radius) of the cycle.

• There might be a subtle tendency for red crosses to be slightly closer to the center than black dots, suggesting mild underprediction or damping in the simulated model.

💡 When to Use:

Use this plot as a primary diagnostic tool to:

• Get an initial visual assessment of point-forecast accuracy.

• Quickly identify overall model bias (systematic over/under prediction).

• Spot specific samples or regions (if angle is meaningful) with large prediction errors.

• Complement numerical scores (MAE, RMSE) with an intuitive overview of model fit, especially for cyclical or ordered data.

---

Anomaly Magnitude

Highlights instances where the actual value falls outside the prediction interval [Qlow, Qup]. Shows the location (angle), type (color), and severity (radius) of anomalies.

import kdiagram as kd
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt

# --- Data Generation ---
np.random.seed(42)
n_points = 180
df = pd.DataFrame({'sample_id': range(n_points)})
df['actual'] = np.random.normal(loc=20, scale=5, size=n_points)
df['q10'] = df['actual'] - np.random.uniform(2, 6, size=n_points)
df['q90'] = df['actual'] + np.random.uniform(2, 6, size=n_points)
# Add anomalies
under_indices = np.random.choice(n_points, 20, replace=False)
df.loc[under_indices, 'actual'] = df.loc[under_indices, 'q10'] - \
                                   np.random.uniform(1, 5, size=20)
available = list(set(range(n_points)) - set(under_indices))
over_indices = np.random.choice(available, 20, replace=False)
df.loc[over_indices, 'actual'] = df.loc[over_indices, 'q90'] + \
                                  np.random.uniform(1, 5, size=20)

# --- Plotting ---
kd.plot_anomaly_magnitude(
    df=df,
    actual_col='actual',
    q_cols=['q10', 'q90'],
    title="Gallery: Prediction Anomaly Magnitude",
    cbar=True,
    s=30,
    verbose=0, # Keep output clean for gallery
    # Save the plot (adjust path relative to docs/source/)
    savefig="gallery/images/gallery_anomaly_magnitude.png"
)
plt.close()

🧠 Analysis and Interpretation

The Anomaly Magnitude Plot provides valuable insights into prediction interval failures, showing how far actual values deviate when they fall outside the predicted bounds (e.g., Q10 and Q90). Only points representing anomalies are plotted.

Key Features:

• Angle (θ): Represents the sample’s position or index in the dataset, arranged circularly.

• Radius (r): Directly corresponds to the magnitude of the anomaly (\(|y_{actual} - y_{bound}|\)). Larger radii indicate more severe prediction interval failures.

• Color: Distinguishes the type of anomaly using different colormaps (defaults: Blues for under-prediction, Reds for over-prediction).

• Color Intensity: Further emphasizes the anomaly’s severity, with darker/more intense colors typically representing larger magnitudes (larger radius).

🔍 In this Example:

• The plot clearly separates under-predictions (blue points, where actual < q10) and over-predictions (red points, where actual > q90).

• Points further from the center represent larger deviations from the predicted [q10, q90] range. We can visually identify the most significant prediction failures.

• The angular distribution shows where these failures occur within the sample order. Clusters might indicate problematic regimes.

💡 When to Use:

This plot is essential for diagnosing model uncertainty calibration and identifying high-risk predictions:

• Pinpoint Interval Failures: Identify exactly which samples fall outside the expected range.

• Assess Anomaly Severity: Quantify how far outside the bounds the actual values lie.

• Analyze Error Type: Determine if the model tends to fail more often through under-prediction or over-prediction.

• Guide Model Refinement: Focus attention on samples/regions with large anomalies where uncertainty estimation needs improvement.

It offers a geographically or temporally focused investigation into where and how prediction intervals fail, complementing plots that assess point forecast accuracy.

---

Overall Coverage

Calculates and displays the overall empirical coverage rate(s) compared to the nominal rate. Useful for comparing average interval calibration across models. Shown here with a radar plot for two simulated models.

import kdiagram as kd
import numpy as np
import matplotlib.pyplot as plt

# --- Data Generation ---
np.random.seed(42)
y_true = np.random.rand(100) * 10
# Model 1 (e.g., ~80% coverage)
y_pred_q1 = np.sort(np.random.normal(
    loc=y_true[:, np.newaxis], scale=1.5, size=(100, 2)), axis=1)
# Model 2 (e.g., ~60% coverage - narrower intervals)
y_pred_q2 = np.sort(np.random.normal(
    loc=y_true[:, np.newaxis], scale=0.8, size=(100, 2)), axis=1)
q_levels = [0.1, 0.9] # Nominal 80% interval

# --- Plotting ---
kd.plot_coverage(
    y_true,
    y_pred_q1,
    y_pred_q2,
    names=['Model A (Wider)', 'Model B (Narrower)'],
    q=q_levels,
    kind='radar', # Use radar chart for profile comparison
    title='Gallery: Overall Coverage Comparison (Radar)',
    cov_fill=True,
    verbose=0,
    # Save the plot (adjust path relative to docs/source/)
    savefig="gallery/images/gallery_coverage_radar.png"
)
plt.close()

Analysis and Interpretation

This plot compares the overall empirical coverage rate between two simulated models using a radar plot. It helps assess the interval calibration across models, evaluating how well their predicted intervals (e.g., Q10 to Q90, implying 80% nominal coverage here) capture the actual values on average over the dataset.

Analysis and Interpretation:

In this example radar plot:

• Model A (Wider): Exhibits a higher coverage rate

  (closer to the outer edge, likely near the target 80% or higher). This indicates its wider prediction intervals successfully encompass a larger fraction of the true values. While seemingly safer, it might suggest the model is conservative, potentially overestimating uncertainty.

• Model B (Narrower): Shows a lower coverage rate (points

  closer to the center). Its narrower intervals fail to capture the true value more often. This model might seem more precise but likely underestimates uncertainty, increasing the risk of errors where reality falls outside the predicted range.

The radar layout effectively contrasts the coverage profiles. Points closer to the outer boundary (radius 1.0) represent better average coverage relative to the defined interval.

When to Use This Plot:

• Comparing Interval Calibration: Ideal for a high-level

  comparison of how well different models’ uncertainty estimates are calibrated (on average). Is one model consistently too wide (over-covered) or too narrow (under-covered)?

• Model Selection: Aids in selecting a model based on risk

  tolerance. Model A might be preferred for risk-averse tasks, while Model B might be chosen if tighter (though less reliable) intervals are desired.

• Summarizing Reliability: Provides a concise summary of the

  average reliability of prediction intervals.

---

Coverage Diagnostic

Visualizes coverage success (radius 1) or failure (radius 0) for each individual data point. Helps diagnose where intervals fail. The solid line shows the overall average coverage rate. Shown here using bars.

import kdiagram as kd
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt

# --- Data Generation ---
np.random.seed(88)
n_points = 200
df = pd.DataFrame({'point_id': range(n_points)})
df['actual_val'] = np.random.normal(loc=5, scale=1.5, size=n_points)
df['q_lower'] = 5 - np.random.uniform(1, 3, n_points)
df['q_upper'] = 5 + np.random.uniform(1, 3, n_points)
# Some points deliberately outside
df.loc[::15, 'actual_val'] = df.loc[::15, 'q_upper'] + 1

# --- Plotting ---
kd.plot_coverage_diagnostic(
    df=df,
    actual_col='actual_val',
    q_cols=['q_lower', 'q_upper'],
    title='Gallery: Point-wise Coverage Diagnostic (Bars)',
    as_bars=True, # Display as bars instead of scatter
    fill_gradient=True, # Show background gradient
    coverage_line_color='darkorange', # Example customization
    verbose=0,
    # Save the plot (adjust path relative to docs/source/)
    savefig="gallery/images/gallery_coverage_diagnostic_bars.png"
)
plt.close()

🧠 Analysis and Interpretation

This plot provides a point-wise coverage diagnostic, showing if the actual value for each sample falls within the prediction interval (e.g., Q10-Q90). Each bar (or point if as_bars=False) represents one sample, arranged angularly by index.

🔍 Key Insights from this Example:

• Bar Height/Radius: Indicates coverage status. A bar

  reaching radius 1 means the actual value was inside the interval (success). A bar at radius 0 means the actual value was outside (failure).

• Color (Implied): Although not the primary focus here,

  the points/bars are often colored by coverage status (e.g., using the cmap parameter, green for 1, red for 0).

• Average Coverage Line: The solid circular line (orange

  in this example code’s customization) is drawn at the radius corresponding to the overall coverage rate (e.g., 0.75 if 75% of points are covered). This provides an immediate visual benchmark against the nominal target (e.g., 0.80 for a Q10-Q90 interval) and the plot boundaries (0 & 1).

• Patterns: Look for clusters of bars at radius 0. These

  indicate ranges of samples (or specific conditions if the angle represented something else) where the model’s intervals consistently fail.

💡 When to Use This Plot:

• Diagnosing Interval Failures: Go beyond the average score

  provided by plot_coverage to see which specific samples are missed by the prediction intervals.

• Identifying Systematic Errors: Determine if coverage

  failures are random or concentrated in certain parts of the data distribution (represented by angles).

• Visual Calibration Assessment: Get a detailed view of how

  well the empirical coverage matches the nominal rate point- by-point, complementing the overall average line.

• Guiding Model Improvement: Pinpoint problematic samples

  or regimes where uncertainty quantification needs refinement.

---

Interval Consistency

Analyzes the stability of the prediction interval width (Qup - Qlow) for each location over multiple time steps. Radius shows variability (CV or Std Dev); color often shows average Q50. High radius means inconsistent width.

import kdiagram as kd
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt

# --- Data Generation ---
np.random.seed(42)
n_points = 100
n_years = 4
years = list(range(2021, 2021 + n_years))
df = pd.DataFrame({'id': range(n_points)})
qlow_cols, qup_cols, q50_cols = [], [], []
for i, year in enumerate(years):
    ql, qu, q50 = f'val_{year}_q10', f'val_{year}_q90', f'val_{year}_q50'
    qlow_cols.append(ql); qup_cols.append(qu); q50_cols.append(q50)
    base_low = np.random.rand(n_points)*5 + i*0.2
    width = np.random.rand(n_points)*3 + 1 + np.sin(
        np.linspace(0, np.pi, n_points))*i # Vary width
    df[ql] = base_low; df[qu] = base_low + width
    df[q50] = base_low + width/2 + np.random.randn(n_points)*0.5

# --- Plotting ---
kd.plot_interval_consistency(
    df=df,
    qlow_cols=qlow_cols,
    qup_cols=qup_cols,
    q50_cols=q50_cols, # Color by average Q50
    use_cv=True,       # Radius = Coefficient of Variation of width
    title='Gallery: Interval Width Consistency (CV)',
    acov='half_circle',
    cmap='viridis',
    # Save the plot (adjust path relative to docs/source/)
    savefig="gallery/images/gallery_interval_consistency_cv.png"
)
plt.close()

🧠 Analysis and Interpretation

This plot analyzes the stability of prediction interval widths (e.g., Q90 - Q10) over multiple time steps or forecast horizons for different samples (locations/indices arranged angularly).

Key Features:

• Radius (r): Corresponds to the variability of the interval width over time for each sample. By default (use_cv=True), it shows the Coefficient of Variation (CV), representing relative variability. If use_cv=False, it shows the standard deviation (absolute variability). * Large Radius: High inconsistency (width fluctuates a lot). * Small Radius: High consistency (width is stable).

• Color: Typically represents the average Q50 (median prediction) across the time steps for each sample, providing context about the prediction magnitude. Darker/cooler colors often indicate lower average Q50, brighter/warmer colors indicate higher average Q50 (depending on the cmap).

• Angle (θ): Represents the sample index or location.

🔍 Key Insights from this Example:

• Points far from the center indicate locations where the model’s uncertainty estimate (interval width) is less stable across the different years included in the data.

• Points clustered near the center represent locations with consistent interval widths over time.

• The color mapping (using viridis) shows whether high/low consistency (radius) correlates with high/low average predicted values (color). For instance, are the most inconsistent predictions (large radius) happening in areas predicted to have high values (yellow) or low values (purple)?

💡 When to Use This Plot:

• Assess Model Stability: Identify samples/locations where uncertainty predictions are erratic or stable over time/horizons.

• Diagnose Uncertainty Drift: While other plots show average drift, this shows the variability aspect of drift for each point.

• Compare Relative vs. Absolute Variability: Toggle use_cv to understand if large fluctuations are significant relative to the mean width (CV) or just large in absolute terms (Std Dev).

• Guide Risk Assessment: Focus on predictions where interval widths are stable (low radius) for more reliable planning, and treat predictions with high variability (high radius) with more caution.

---

Interval Width

Visualizes the magnitude of the prediction interval width (Qup - Qlow) for each sample at a single time point. Radius directly represents the width. Color can represent width or an optional third variable (z_col), here showing the Q50 prediction.

import kdiagram as kd
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt

# --- Data Generation ---
np.random.seed(77)
n_points = 150
df = pd.DataFrame({'location': range(n_points)})
df['elevation'] = np.linspace(100, 500, n_points) # Example feature
df['q10_val'] = np.random.rand(n_points) * 20
# Width depends on elevation in this synthetic example
width = 5 + (df['elevation'] / 100) * np.random.uniform(0.5, 2, n_points)
df['q90_val'] = df['q10_val'] + width
df['q50_val'] = df['q10_val'] + width / 2 # Use as z_col

# --- Plotting ---
kd.plot_interval_width(
    df=df,
    q_cols=['q10_val', 'q90_val'],
    z_col='q50_val', # Color points by Q50 value
    title='Gallery: Interval Width (Colored by Q50)',
    cmap='plasma',
    cbar=True,
    s=30,
    # Save the plot (adjust path relative to docs/source/)
    savefig="gallery/images/gallery_interval_width_z.png"
)
plt.close()

🧠 Analysis and Interpretation

This plot shows the magnitude of predicted uncertainty, represented by the interval width (e.g., Q90 - Q10), for each sample at a specific time point or forecast horizon.

Key Features:

• Radius (r): Directly proportional to the interval width. Larger radius means greater predicted uncertainty for that sample.

• Angle (θ): Represents the sample index or location, arranged circularly.

• Color: Represents the value of the column specified by the z_col parameter (here, the Q50 median prediction). If z_col is not provided, color defaults to representing the interval width (radius).

🔍 Key Insights from this Example:

• We can visually identify samples with the widest (points furthest from center) and narrowest (points closest to center) prediction intervals.

• The plasma colormap colors points by their Q50 value (yellow = high Q50, purple = low Q50). By combining radius and color, we can assess if higher uncertainty (larger radius) tends to occur for samples with higher or lower median predictions (color). In this synthetic example, width was linked to ‘elevation’, which might also correlate with Q50, potentially revealing a pattern.

💡 When to Use This Plot:

• Visualize Uncertainty Magnitude: Get a direct overview of how much uncertainty the model predicts for each sample.

• Identify High/Low Uncertainty Samples: Quickly spot the most and least certain predictions.

• Explore Correlations: Use the z_col parameter to investigate if uncertainty width correlates with other factors like the magnitude of the prediction itself (Q50), actual values, or input features.

• Assess Spatial Patterns: If the angle represented spatial location, this plot could reveal geographical areas of high/low predicted uncertainty.

---

Model Drift

Shows how average uncertainty (mean interval width) evolves across different forecast horizons using a polar bar chart. Helps diagnose model degradation over lead time.

import kdiagram as kd
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt

# --- Data Generation ---
np.random.seed(0)
years = [2023, 2024, 2025, 2026, 2027]
n_samples = 50
df = pd.DataFrame()
q10_cols, q90_cols = [], []
for i, year in enumerate(years):
    ql, qu = f'val_{year}_q10', f'val_{year}_q90'
    q10_cols.append(ql); q90_cols.append(qu)
    q10 = np.random.rand(n_samples)*5 + i*0.5 # Width tends to increase
    q90 = q10 + np.random.rand(n_samples)*2 + 1 + i*0.8
    df[ql]=q10; df[qu]=q90

# --- Plotting ---
kd.plot_model_drift(
    df=df,
    q10_cols=q10_cols,
    q90_cols=q90_cols,
    horizons=years, # Label bars with years
    acov='quarter_circle', # Use 90 degree span
    title='Gallery: Model Drift Across Horizons',
    # Save the plot (adjust path relative to docs/source/)
    savefig="gallery/images/gallery_model_drift.png"
)
plt.close()

🧠 Analysis and Interpretation

This Model Drift plot uses a polar bar chart to visualize how the average uncertainty (mean interval width, Q90-Q10) evolves across different forecast horizons (years in this case, arranged angularly).

Analysis and Interpretation:

• Radius (Avg. Uncertainty Width): The length of each bar (its radius) directly represents the average width of the prediction intervals for that specific horizon. Longer bars mean wider average intervals and thus higher average uncertainty for that forecast lead time.

• Angle (Horizon): Each bar corresponds to a successive forecast horizon (e.g., 2023, 2024,…), arranged around the circle.

• Color Gradient: The color often transitions (e.g., cool to warm colors via the default coolwarm cmap) along the angular axis, visually reinforcing the progression through forecast horizons.

🔍 Key Insights from this Example:

• The bars increase in length as we move from earlier years (e.g., 2023) to later years (e.g., 2027) along the angular axis. This clearly indicates model drift: the average uncertainty grows as the forecast horizon extends further into the future.

• The color transition from blue/green towards red mirrors this increase in uncertainty over time.

• This pattern is typical in forecasting, reflecting the increasing difficulty and accumulated error when predicting further ahead. The plot helps quantify this degradation rate.

💡 When to Use This Plot:

• Assess Uncertainty Evolution: Evaluate if and how quickly average forecast uncertainty increases with lead time.

• Monitor Model Degradation: Identify horizons where the uncertainty becomes unacceptably large, indicating the limits of the model’s reliable forecast range.

• Inform Retraining/Updates: Significant drift can signal the need to retrain the model more frequently or incorporate time-dependent features.

• Communicate Forecast Reliability: Show stakeholders how confidence in forecasts typically decreases for longer lead times.

---

Temporal Uncertainty

A general polar scatter plot for visualizing multiple data series. Often used to show different quantiles (e.g., Q10, Q50, Q90) for a single time step to illustrate the uncertainty spread across samples.

import kdiagram as kd
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt

# --- Data Generation ---
np.random.seed(99)
n_points = 80
df = pd.DataFrame({'id': range(n_points)})
base = 10 + 5*np.sin(np.linspace(0, 2*np.pi, n_points))
df['val_q10'] = base - np.random.rand(n_points)*2 - 1
df['val_q50'] = base + np.random.randn(n_points)*0.5
df['val_q90'] = base + np.random.rand(n_points)*2 + 1
# Ensure order for clarity in plot
df['val_q50'] = np.maximum(df['val_q10'] + 0.1, df['val_q50'])
df['val_q90'] = np.maximum(df['val_q50'] + 0.1, df['val_q90'])


# --- Plotting ---
kd.plot_temporal_uncertainty(
    df=df,
    q_cols=['val_q10', 'val_q50', 'val_q90'],
    names=['Q10', 'Q50', 'Q90'],
    title='Gallery: Uncertainty Spread (Q10, Q50, Q90)',
    normalize=False, # Show raw values
    cmap='coolwarm', # Use diverging map for bounds
    s=20,
    mask_angle=True,
    # Save the plot (adjust path relative to docs/source/)
    savefig="gallery/images/gallery_temporal_uncertainty_quantiles.png"
)
plt.close()

🧠 Analysis and Interpretation

This plot uses a polar scatter format to visualize the spread of uncertainty at a single time point by plotting multiple related series, typically different quantile predictions (like Q10, Q50, Q90 shown here).

Analysis and Interpretation:

• Angle (θ): Each angular position represents a unique sample or location from the dataset (ordered by index here).

• Radius (r): The distance from the center represents the actual predicted value for a specific quantile at that sample (since normalize=False was used).

• Color: Each quantile series (Q10, Q50, Q90) is assigned a distinct color (using the coolwarm cmap here, blue for Q10, red for Q90), allowing visual differentiation.

🔍 Key Insights from this Example:

• The radial distance between the blue (Q10) and red (Q90) points at any given angle visually represents the prediction interval width (uncertainty magnitude) for that specific sample.

• We can see how this spread varies around the circle. Some samples (angles) have a larger gap between blue and red points (higher uncertainty), while others have a smaller gap (lower uncertainty).

• The grey points (Q50, median) trace the central tendency, lying between the Q10 and Q90 bounds.

• The overall pattern follows the sinusoidal base signal used in the data generation.

💡 When to Use This Plot:

• Visualize Interval Spread: Show the range between lower and upper quantile bounds for each sample simultaneously at a specific time/horizon.

• Compare Multiple Series: Plot predictions from different models side-by-side against the same angular axis.

• Identify Uncertainty Patterns: See if uncertainty (spread between quantiles) correlates with sample index or location (angle) or with the magnitude of the prediction (radius/color).

• Check Quantile Ordering: Visually verify that Q10 <= Q50 <= Q90 holds for most samples.

---

Uncertainty Drift

Visualizes how the interval width pattern evolves across multiple time steps using concentric rings. Each ring represents a time step, showing the relative uncertainty width at each angle (location).

import kdiagram as kd
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt

# --- Data Generation ---
np.random.seed(55)
n_points = 90; n_years = 4; years = range(2020, 2020 + n_years)
df = pd.DataFrame({'id': range(n_points)})
qlow_cols, qup_cols = [], []
for i, year in enumerate(years):
    ql, qu = f'value_{year}_q10', f'value_{year}_q90'
    qlow_cols.append(ql); qup_cols.append(qu)
    base_low = np.random.rand(n_points)*3 + i*0.1
    width = (np.random.rand(n_points)+0.5)*(1.5+i*0.3 + np.cos(
        np.linspace(0, 2*np.pi, n_points)))
    df[ql] = base_low; df[qu] = base_low + width
    df[qu] = np.maximum(df[qu], df[ql]) # Ensure non-negative width

# --- Plotting ---
kd.plot_uncertainty_drift(
    df=df,
    qlow_cols=qlow_cols,
    qup_cols=qup_cols,
    dt_labels=[str(y) for y in years],
    title='Gallery: Uncertainty Drift (Rings)',
    cmap='magma',
    base_radius=0.1, band_height=0.1,
    # Save the plot (adjust path relative to docs/source/)
    savefig="gallery/images/gallery_uncertainty_drift_rings.png"
)
plt.close()

🧠 Analysis and Interpretation

This plot displays how the prediction interval width pattern (Q90-Q10) changes over multiple time steps (e.g., years) using concentric rings. Each ring represents a specific time step, ordered radially outwards.

Analysis and Interpretation:

• Rings & Time: Each colored ring corresponds to a time step (e.g., 2020 near center, 2023 further out). The legend links colors to time steps.

• Radius & Uncertainty: The radius of a point on a specific ring represents the relative interval width for that sample at that time. The radius is calculated as a base offset for the ring plus a component scaled by the globally normalized width. Therefore, bulges or larger radii on a ring indicate higher relative uncertainty for those samples at that time.

• Comparing Rings: Observe how the overall size and shape of the rings change from inner (earlier) to outer (later). Increasing average radius or increased “bumpiness” in outer rings suggests uncertainty drift - uncertainty grows or becomes more variable over time.

• Angular Patterns: Consistent high/low radii at specific angles across multiple rings pinpoint locations/samples with persistently high/low relative uncertainty.

🔍 Key Insights from this Example:

• The concentric rings clearly separate the uncertainty patterns for different years (2020-2023).

• Comparing the rings reveals how the spatial distribution and magnitude of relative uncertainty change over the forecast horizon. For instance, one might observe uncertainty increasing overall (outer rings generally larger) or becoming more pronounced in certain angular sectors (locations).

• Potential cyclic patterns in width along the angular axis might suggest seasonal or location-based effects on uncertainty.

💡 When to Use This Plot:

• Visualize Uncertainty Evolution: Track how the entire pattern of uncertainty changes across multiple forecast periods.

• Identify Temporal Drift Patterns: See if uncertainty increases uniformly, or only in specific regions/samples over time.

• Compare Uncertainty Maps: Overlay and compare the “uncertainty map” (relative interval width vs. sample index/location) from different time steps in a single view.

• Assess Long-Term Reliability: Evaluate if the model’s uncertainty estimates remain stable or degrade significantly as forecasts extend further out.

---

Prediction Velocity

Visualizes the average rate of change (velocity) of the median (Q50) prediction over consecutive time periods for each location. Radius indicates velocity magnitude; color can indicate velocity or average Q50.

import kdiagram as kd
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt

# --- Data Generation ---
np.random.seed(123)
n_points = 100; years = range(2020, 2024)
df = pd.DataFrame({'location_id': range(n_points)})
q50_cols = []
base_val = np.random.rand(n_points)*10
trend = np.linspace(0, 5, n_points)
for i, year in enumerate(years):
    q50_col = f'val_{year}_q50'
    q50_cols.append(q50_col)
    noise = np.random.randn(n_points)*0.5
    df[q50_col] = base_val + trend*i + noise

# --- Plotting ---
kd.plot_velocity(
    df=df,
    q50_cols=q50_cols,
    title='Gallery: Prediction Velocity (Colored by Avg Q50)',
    use_abs_color=True, # Color by magnitude of Q50
    normalize=True,     # Normalize radius (velocity)
    cmap='cividis',
    cbar=True,
    s=25,
    # Save the plot (adjust path relative to docs/source/)
    savefig="gallery/images/gallery_velocity_abs_color.png"
)
plt.close()

🧠 Analysis and Interpretation

This plot visualizes the average rate of change (velocity) of the median (Q50) prediction across consecutive time periods. Each point represents a sample/location.

Analysis and Interpretation:

• Radius (Velocity Magnitude): The distance from the center indicates the average speed at which the Q50 prediction is changing over time for that sample. Larger radii mean faster average change (positive or negative); smaller radii mean more stable Q50 predictions. (Note: If normalize=False, radius shows raw velocity).

• Angle (θ): Represents the sample index/location, arranged circularly.

• Color (Context): The color provides context.

  • If use_abs_color=True (default, as in this example): Color maps to the average absolute Q50 value across periods. This helps see if rapid changes (high radius) occur in high-value (e.g., yellow in cividis) or low-value (e.g., purple) regions.

  • If use_abs_color=False: Color maps directly to the velocity value. Using a diverging colormap (like ‘coolwarm’) distinguishes between positive velocity (increasing trend) and negative velocity (decreasing trend).

🔍 Key Insights from this Example:

• Points far from the center highlight locations where the median prediction is changing most rapidly on average between the years provided.

• The cividis colormap shows the average magnitude of the Q50 prediction at each location. We can observe if the high-velocity points (large radius) coincide with high-magnitude (yellow) or low-magnitude (purple) predictions.

• Clustering of points with similar radius/color might indicate spatial patterns in the phenomenon’s dynamics.

💡 When to Use This Plot:

• Identify Dynamic Hotspots: Find samples/locations where the central forecast trend is changing most quickly.

• Assess Prediction Stability: Locate areas where predictions are relatively stable (low velocity) vs. dynamic (high velocity).

• Contextualize Change Rate: Use use_abs_color=True to see if rapid changes are happening in already critical high/low value areas. Use use_abs_color=False with a diverging map to see the direction (increase/decrease) of the average change.

• Analyze Temporal Trends Spatially: Understand the spatial distribution of the rate of change across different locations.

---