[BUG] Negative loss amounts in BootstrapODPSample

[cmds2k]

BootstrapODPSample creates samples with negative loss amounts. This leads to negative ldfs, cdfs and ultimates. There should be a way to keep samples with positive losses only.

[kennethshsu]

Hi @cmds2k, can you share some code examples?

[ntwong319]

Hi @kennethshsu I wanted to bring up this bug. I was running an ODP bootstrapping on our dataset and ran into the same issue OP posted about. Here are some snippets of my code:

samples_cdf = cl.Development(drop_high=True, drop_low=True,average='volume').fit(samples).cdf_

pipe = cl.Pipeline(
    [('dev', cl.Development(drop_high=True, drop_low=True,average='volume')),
     ('tail', cl.TailConstant(1.00)),
     ('model', cl.Chainladder())]).fit(samples)

resampled_ldf = pipe.named_steps.dev.ldf_

from scipy.stats.mstats import winsorize

fig, axes = plt.subplots(4, 4, figsize=(30, 12))
axes = axes.flatten()
#percentile_50_values = []

for x in range(16):
    # Apply winsorization to your data
    winsorized_data = winsorize(pd.Series(resampled_ldf.values[:, 0, 0, x]), limits=(0.01, 0.01))  # Adjust the limits as needed

    # Plot the histogram with winsorized data
    pd.Series(winsorized_data).plot(kind='hist', bins=100, ax=axes[x])
    axes[x].set_title(f'Age {(x+1)*3}-{(x+1)*3+3} LDF distribution using Bootstrap')
    axes[x].axvline(orig_dev.values[0, 0, 0, x], color="black", linestyle="dashed")

    # Calculate and display the 75th percentile
    percentile_75 = np.percentile(winsorized_data, 75)
    axes[x].text(0.6, 0.7, f"75th Percentile: {percentile_75:.3f}", transform=axes[x].transAxes)

    percentile_50 = np.percentile(winsorized_data, 50)
    axes[x].text(0.6, 0.8, f"50th Percentile: {percentile_50:.3f}", transform=axes[x].transAxes)

    # Display the original value
    axes[x].text(0.6, 0.9, f"Original LDF: {round(orig_dev.values[0, 0, 0, x], 4)}", transform=axes[x].transAxes)

    percentile_50 = np.percentile(winsorized_data, 50)
    

fig.suptitle("LDFs Using ODP Bootstrap")

# Add a common y-label for all subplots
fig.text(0.5, 0.04, 'Value', ha='center')
plt.subplots_adjust(wspace=0.3, hspace=0.3)

# Show the figure
plt.show()

And it generated this LDF distribution plot for the 24-27 factor.

Additionally, I wanted to see what the minimum IBNR produced by this command was:

pipe.predict(net["REP_LALAE"]).ibnr_.min()

And the output was