yatshunlee / caviar Goto Github PK

Initial guess: random start may not be good enough for that

CaviarModel(0.05, model='adaptive', method='mle')

trial 1: beta = np.array([-0.64774347])

trial 2: beta = np.array([-84.26894253])
show error: D:\CAViaR-Project\test..\caviar_caviar_function.py:20: RuntimeWarning: overflow encountered in exp
1 / (1 + np.exp(G * (returns[t - 1] - sigmas[t - 1]))) - quantile)

CaviarModel(0.05, model='asymmetric', method='mle')

trial 1: np.array([-3.38156961, -0.76269892, -0.0212416 , 0.90496001])

trial 2: beta = np.array([-0.0308821 , 0.98321657, 0.10719172, 1. ])

trial 3: beta = np.array([-7.38144696e-01, 6.16001322e-01, 3.56295801e-09, 5.42463432e-01])

CaviarModel(0.05, model='symmetric', method='mle')

trial 1: beta = np.array([-0.04064223, 0.87657701, -0.23026562])
trial 2: beta = np.array([-0.04058351, 0.87658738, -0.23027769])
trial 4: beta = np.array([-0.04076704, 0.87653064, -0.23027652])

trial 3: beta = np.array([-0.80190414, 0.43086385, -0.30877964])

CaviarModel(0.05, model='igarch', method='mle')

trial 1: beta = np.array([1.65923914, 0.51150889, 0.70126179])

trial 2: beta = np.array([1.14950421, 0.47399757, 0.57122966])

They all showed inconsistence in the convergence.

https://stackoverflow.com/questions/75717959/quantile-regression-unable-to-regress-when-quantile-0-5

Case:

adaptive numeric 0.01 JPM
when m = 1
Update 0: 0.10992994419739356
Update 1: 0.10992874001704382
Update 2: 0.10992873996116201
when m = 2
Update 0: 0.10993328508627404
Update 1: 0.10993022133909976
Update 2: 0.10993022133909976
when m = 3
Update 0: 0.10993364587662575
Update 1: 0.10995754038106954
Final loss: 0.10992873996116201

Training

Testing