218
edits
Changes
Jump to navigation
Jump to search
'''empirics_match_specific_1st_stage_cmaes''':'''empirics_match_specific_2nd_stage_cmaesempirics_match_specific_with_se''':'''empirics_match_specific_with_se''': Saved if task = 'data'
Estimating Unobserved Complementarities between Entrepreneurs and Venture Capitalists Matlab Code (view source)
Revision as of 16:03, 11 June 2018
, 16:03, 11 June 2018→Important contractions
Main Project here: [[Estimating Unobserved Complementarities between Entrepreneurs and Venture Capitalists]]
==Important contractions==
*GMM: Generalized Method of Moments
*MSM: Method of Simulated Moments
*MLE: Maximum Likelihood Estimation
*GA: Genetic Algorithm
*SE: Standard Error
==How to run==
First, apparently, you need a gurobi license, which can be obtained (free) [http://www.gurobi.com/registration/academic-license-reg here].
In master.m, edit the options section to reflect what you want the code to do. Then run it.
'''task:''' Can take the values {'data', 'monte', 'monte_data'}.
'''estimator:''' Can take the values {'MLE', 'MSM', 'compare'}.(other estimators removed in readjusted code)* MSM stands for Markov State Model* MLE stands for Maximum Likelihood Estimation* compare** compare is valid only under task='monte''''use_solver:''' Can take the values {'fminunc', 'ga', 'patternsearch', 'cmaes'}.(other solvers removed in readjusted code)* GA stands for genetic algorithm* fminunc* patternsearch* cmaes'''error_type''' (currently hard coded as 1and isn't fully written to support 2): 1 for match specific errors, with the error distribution following an exchangeable structure. var(e) = sig^2, and cov(e,e') = 1/4*sig^2. 2 for agent specific errors, with the error structure of match <i, j> as sig*ei*ej. ===gmm_2stage_estimation.m=== Does the majority of the work for this problem. Runs the GA, saves the results to 'empirics_match_specific_1st_stage_ga', then runs it again with different globals, and saves it to 'empirics_match_specific_2nd_stage_ga'. ===nonlinearcons_msm.m=== Constraints on GA. For [c,ceq] = nonlinearcons_msm(x), GA constrains x such that c ≤ 0 and ceq = 0. c and ceq are row vectors when there are multiple constraints. ceq is unused for our purposes. ===msmf_corr_coeff.m=== This is the fitness function. Takes a vector and returns a scalar. GA minimizes this function.
===strip_mastermoments.m===Seems to be a version of master.m that only uses Generates the parameters task = 'monte_data', use_solver = 'ga', and estimator = 'MSM' (which apparently are moments needed for the only values that work for master.m, but unverified)GMM.
==Location/Structure of Data==
'''empirics_match_specific_1st_stage_ga''':
'''empirics_match_specific_2nd_stage_ga''':
'''type_(error_type)_monte_mktsize_(mktsize)_K_(K)_S_(S)_rep_(montenn)''':
==Bugs==
Simulated moments matrix dimension mismatch. This bug (also in the adjusted code) does not allow the "optimal stage" to complete.
Assignment has more non-singleton rhs dimensions than non-singleton subscripts
Error in gmm_2stage_estimation (line 65)
EV(:, :, m) = EV(:, :, m) + temp2;
Error in master (line 513)
gmm_2stage_estimation;
This is caused because temp2 is 166x166 in monte_data, but nm = M = 1, so EV is 1x1. nm is set to size(M0, 1), but M0 is sometimes a scalar?
This is solved by commenting out mkt_resample.m line 107:
% M0 = moments(M, 1, S, K, pv', asg, H, HK, G, FF, Hu, Hd, z1, z2, psa, ps1, ps2, N1, N2)
because it returns a scalar, but should return a 166x1 vector.
Better fix notes:
The problem is narrowed down to the section of mkt_resample.m before the call to moments. Somehow the parameters to moments are changed such that it returns a scalar.
Solution found (I think):
monte_M cannot be set to 1. For values greater than 1, the program seems to work. It doesn't crash, anyway.
==Profiling==
Ran first iteration with task='monte', use_solver='ga', estimator='MSM', gurobi output disabled (took 42 sec to finish 1st iteration)
[[File:Unobserved comp profile monte.png]]
Monte_data and monte run much faster than data, because data is so large.
==Notes==
There is a lot of unused code. For example, monte_std and monte_x seem to only be assigned, not used.
