'n_periods' does not respect dropped link ratios

[andrewwilson201]

Are you on the latest chainladder version?

• Yes, this bug occurs on the latest version.

Describe the bug in words

Description

Currently, when using the Development estimator, the n_periods parameter is applied before any exclusions from drop, drop_valuation, etc., are considered. This leads to unexpected results when attempting to exclude specific origin-development combinations from LDF calculations.

Current Behavior

In the example below when n_periods=1 is specified along with drop=[("2009", 12)], the system:

1. Identifies the most recent period (2009)
2. Applies the drop, resulting in no valid periods for the calculation
3. Returns NaN or uses an incorrect period

Expected Behavior

When n_periods=1 is specified along with drop=[("2009", 12)], the system should:

1. Identify which periods are valid (non-dropped)
2. Select the most recent valid period (skipping 2009)
3. Use that period (e.g., 2008) for the LDF calculation

How can the bug be reproduced?

import chainladder as cl
import pandas as pd

# create example triangle
data = {
    'origin': ["2007-01-01", "2007-01-01", "2007-01-01", "2007-01-01", "2008-01-01", "2008-01-01", "2008-01-01", "2009-01-01", "2009-01-01", "2010-01-01"],
    'development': ["2007-01-01", "2008-01-01", "2009-01-01", "2010-01-01", "2008-01-01", "2009-01-01", "2010-01-01", "2009-01-01", "2010-01-01", "2010-01-01"],
    'loss': [100, 200, 300, 400, 150, 300, 450, 200, 250, 50]
}

df = pd.DataFrame(data)
tri = cl.Triangle(
    df, 
    origin='origin',
    development='development',
    columns='loss',
    cumulative=True
)

# calculate ldf with n_periods=1 and the most recent period (2009) dropped
dev = cl.Development(n_periods=1, drop=[('2009', 12)]).fit(tri)
dev.ldf_

# current behaviour : uses only the excluded ratio from 2009 in dev month 12 resulting in NaN
# desired behaviour: use valid ratios (that haven't been dropped) resulting in the ratio from 2008 in dev month 12 and so expected result is 300 / 150 = 2

What is the expected behavior?

The n_periods parameter should be applied after all drop logic. It should select the n most recent link ratios from the set of all valid, non-dropped, non-NaN data points. This ensures that the n_periods argument is always honored when enough valid data is available.

[kennethshsu]

Hey @andrewwilson201,

When I wrote that piece of the code a long time ago, I debated very hard on the order of operations. And after reading your post, I still think n_periods should be applied first, before other drop options occurs.

The way I think about this is that, supposed we know the upper triangles are no good for the estimation of LDFs, we should discard them first. n_periods is also the weakest argument of all (the most non-specific) right?

I do see the value of having the n_periods as the guard as in use n_periods at minimum, after dropping all illegible factors. Perhaps we make this a user option?

In fact, I think the best solution is that we should not allow the user to give multiple drops at once. It has to be one at a time, and let the user decide the order of operation.

What are your thoughts?

[henrydingliu]

This is semantics. Basically when someone says average 5 excl hi/lo, do they mean take the latest 5 years and exclude hi/lo, or latest 7 years and exclude hi/lo so we are in the end averaging 5 periods. I've seen highly reputable shops in the industry use either. My recommendation is to use the Friedland text as the tie breaker. P87 of Friedland contains the following paragraph

To calculate the reported claims 24-36 month medial average development factor of the latest
5x1, we consider the 24-36 month factors for accident years 2001 through 2005; we exclude the
highest value (1.062 for accident year 2001) and the lowest value (1.055 for accident year 2004)
and take an average of the remaining three values. The 24-36 month medial average of the latest
5x1 is 1.057 ((1.059 + 1.057 + 1.056) / 3).

In this example, Friedland adopts the semantic convention to state the pre-exclusion number of periods, i.e. avg 5 ex HiLo will exclude hilo from the latest 5, ending with 3 periods to average.

The currently implementation in this package faithfully adheres to the Friedland convention. Given our ongoing effort to recreate all Friedland exhibits in order to lower the barrier of adoption for new actuaries, I do not support making any changes.

[kennethshsu]

So if we were to change the existing behavior, we would need to throw deprecation warning first. And I'm not 100% convinced that either the current or proposed solution is better.

I do think that it would be better if we force the user to think about the order of operation. This of course, needs a deprecation warning too. I think this is worth discussing on the call?

[henrydingliu]

I'm 100% convinced that the current implementation is better.

[kennethshsu]

Well thank you lol...

Even if it's not clear to the user the order of the drop operations? Is it not better to require the user to be explicit with the order of the drop operations?

[henrydingliu]

We can add documentation to state that we follow the Friedland order.

[kennethshsu]

@andrewwilson201 I'm interested in hearing from you. Why do you think the current implementation is wrong? Did you reference an example in a paper somewhere? Or was it more like intuition?