Taylor Diagram

How can you summarize multiple aspects of a model’s performance in a single, intuitive plot? The Taylor diagram is a classic solution, brilliantly summarizing how closely a model’s predictions match a reference series by plotting their standard deviation and correlation on a polar axis [1]. This allows for a quick assessment of model fidelity.

The k-diagram library provides three commands for creating them:

• plot-taylor-diagram: For a standard, clean diagram.

• plot-taylor-diagram-in: Adds a colored background for context.

• taylor-diagram: A flexible command that can take either raw data or pre-computed statistics.

Available Commands

Command	Description
plot-taylor-diagram	Generates a standard Taylor Diagram from pre-calculated statistics.
plot-taylor-diagram-in	Generates a Taylor Diagram with background shading from raw data.
taylor-diagram	An alias for the plot-taylor-diagram command.

Common Conventions

All commands on this page read a tabular data file (e.g., data.csv) and require you to specify the ground-truth column with --y-true. You can provide prediction columns using the --pred or the named --model NAME:COL syntax. To save a plot, simply add the --savefig out.png flag.

Note

In a Taylor diagram, the correlation \(\rho\) is mapped to the angle via \(\theta=\arccos(\rho)\), while the model’s standard deviation is mapped to the radius.

plot-taylor-diagram

This is the primary command for creating a standard Taylor diagram. It plots each model as a point and includes a reference arc representing the standard deviation of the true data, making it easy to see which models are closest to the reference.

The command usage is as follows:

k-diagram plot-taylor-diagram INPUT
  --y-true Y_TRUE
  [--pred COL | --model NAME:COL]...
  [--names NAME1 NAME2 ...]
  [--acov half_circle]
  [--zero-location W]
  [--direction -1]

Here’s a typical example comparing two models, “Model A” and “Model B”:

k-diagram plot-taylor-diagram data.csv \
  --y-true y_actual \
  --pred model_a_preds model_b_preds \
  --names "Model A" "Model B" \
  --acov half_circle \
  --savefig taylor_basic.png

plot-taylor-diagram-in

This command enhances the standard diagram by adding a shaded background colormap. The color can represent a diagnostic metric, such as the correlation itself, providing an extra layer of visual context for interpreting the model points.

To generate this plot, you can add background-specific flags:

k-diagram plot-taylor-diagram-in INPUT
  --y-true Y_TRUE
  [--pred COL | --model NAME:COL]...
  [--radial-strategy convergence]
  [--cmap viridis]
  [--cbar]

For example, let’s create a diagram where the background color shows the correlation field:

k-diagram plot-taylor-diagram-in data.csv \
  --y-true y \
  --model A:m1 --model B:m2 \
  --radial-strategy convergence \
  --cmap viridis \
  --cbar \
  --savefig taylor_with_background.png

taylor-diagram

This is a highly flexible command that can operate in two distinct modes, making it useful in a wide variety of situations.

1. Data-mode (from a dataset) This mode works just like the other commands, calculating statistics directly from your data columns.

2. Stats-mode (from pre-computed values) This mode is incredibly useful when you don’t have the raw data but already know the statistics (standard deviation and correlation). It allows you to generate a Taylor diagram without needing an input file.

Here is an example of using stats-mode to plot the performance of three models for which we have pre-computed scores:

k-diagram taylor-diagram \
  --stddev 1.05 0.88 0.75 \
  --corrcoef 0.91 0.72 0.60 \
  --names "Linear Regression" "SVR" "Random Forest" \
  --draw-ref-arc \
  --cmap plasma \
  --radial-strategy rwf \
  --savefig taylor_from_stats.png

—

Troubleshooting & Tips

• Orientation: The diagram’s orientation can be confusing at first. Use the --zero-location (where correlation=1 sits, e.g., ‘E’ for East) and --direction (-1 for clockwise) flags to match your preferred convention.

• Correlation Labels: By default, the angular axis is labeled with correlation values. If you’d rather see degrees, use the --no-angle-to-corr flag.

• Need more help? Run any command with the -h or --help flag to see its full list of options.

• See Also: The Taylor diagram is a great summary tool. For more detailed comparisons, you might use the radar charts in Comparison & Calibration or dive into feature analysis with the tools in Feature-Based Plots.

---

References

[1]

Karl E. Taylor. Summarizing quantitative measures of model performance in a single diagram. Journal of Geophysical Research: Atmospheres, 106(D7):7183–7192, 2001. doi:10.1029/2000JD900719.