kartechbabu / torchblp

pytorch implementation of BLP'95

torchblp

PyTorch implementation of BLP'95.

Acknowledgements

The data comes from:

Andrews, Isaiah; Gentzkow, Matthew; Shapiro, Jesse M., 2017, "Replication Data for: "Measuring the Sensitivity of Parameter Estimates to Estimation Moments"", https://doi.org/10.7910/DVN/LLARSN, Harvard Dataverse, V1

I used the MATLAB code of Andrews, Gentzkow, and Shapiro to construct GMM instruments.

The squarem implementation is adapted from https://github.com/jeffgortmaker/pyblp/tree/master/pyblp.

The code relies on various undocumented tricks to make this work on a GPU.

Usage

You will need Python 3.

Clone the repo, and from within the src directory run:

python torchblp.py  --data-filename ../data/blp95.csv --unobs-filename ../data/blp95unobs.csv 

If you do not have a GPU use the --disable-cuda option. The code should produce identical results, though it will run slower.

Results

With float64 precision, torchblp almost exactly replicates the estimates of Andrews, Gentzkow, and Shapiro (AGS'17). With float32 precision, the torchblp estimates are different from AGS'17 but not by much. The AGS'17 estimates themselves are different from the published BLP'95 estimates for various reasons including a bug in instrument construction in BLP'95 (see the AGS'17 appendix for details). The code takes about 20 secs to run on an NVIDIA Quadro RTX 4000 (about 5 secs are wasted loading and transforming data, the rest is estimation.) Inverting all market shares once takes about 0.12 secs using float64, and 0.07 secs using float32. The code does not produce standard errors, yet.

param_name	BLP'95	AGS'17	torchblp (float64)	torchblp (float32)
alpha_price	43.501	42.870	42.870	43.712
sigma_const	3.612	2.522	2.522	2.108
sigma_hpwt	4.628	3.525	3.525	5.780
sigma_air	1.818	4.167	4.166	3.312
sigma_mpd	1.050	0.393	0.393	0.290
sigma_space	2.056	1.937	1.937	1.823
demand_const	-7.061	-7.728	-7.728	-7.394
demand_hpwt	2.883	4.620	4.620	3.471
demand_air	1.521	-1.227	-1.226	-0.512
demand_mpd	-0.122	0.293	0.293	0.338
demand_space	3.460	3.992	3.992	4.160
supply_const	0.952	2.751	2.751	2.804
supply_loghpwt	0.477	0.812	0.812	0.853
supply_air	0.619	0.430	0.430	0.393
supply_logmpg	-0.415	-0.610	-0.610	-0.617
supply_logspace	-0.046	-0.352	-0.352	-0.362
supply_trend	0.019	0.027	0.027	0.028

References

1. https://academic.oup.com/qje/article/132/4/1553/3861634
2. https://doi.org/10.7910/DVN/LLARSN
3. https://chrisconlon.github.io/site/pyblp.pdf
4. https://github.com/jeffgortmaker/pyblp/tree/master/pyblp