henrydingliu

thanks for the walkthrough! i think i understand a bit better. with that new understanding, ldf_ and cdf_ makes sense to me the way they are specified. I think what...

created #381 for phase 1 fixes. ideally would have liked to create cum_zeta_ and fit_zeta_ as properties. But I think that would only work if I added them directly to...

in further testing, #181 is popping up again. ```python raa = cl.load_sample('raa') raa.values[0,0,0,5:] = 18608 ult = cl.Chainladder().fit(raa) incr = cl.IncrementalAdditive().fit(raa,sample_weight=ult.ultimate_) print(incr.tri_zeta) print(incr.zeta_) ``` Take the 108 column as example....

```python def nan_to_zero(tri,years = None): xp = tri.get_array_module() tric = tri.copy() if years is None: tric.values = xp.nan_to_num(tric.values) * tric.nan_triangle else: years_list = years if isinstance(years,list) else [years] years_np =...

@jbogaardt thoughts?

lemme add a sparse incremental test and tinker around. what's a good sparse data I can use?

based on my primitive understanding of sparse, would it be correct to conclude that [these couple lines](https://github.com/casact/chainladder-python/blob/62bc6952efbc6a7b39d160121523fbf5be909404/chainladder/development/incremental.py#L112) in the class are essentially turning off sparse support? I haven't made full...

I do see how in most cases dense coefficients work just fine. I don't think dense coefficients currently work for incrementaladditive. The conversion from zeta_ to ldf_ is done densely...

when's the cutoff for 0.8,14 @jbogaardt

here is my proposed alternative. basically using selected pattern (not implied) to distribute IBNR. ```python def new_runoff(dev,ultimate): emergence = dev.ldf_.copy() * (ultimate/ultimate) emergence = ( emergence[emergence.valuation < dev.valuation_date] * 0...