calibrated dnn to mkt prices, performance plots

dnn_pricing/dnnMktCalibrate.py

@@ -3,52 +3,84 @@
 from sklearn.metrics import mean_squared_error
 from dnnOptionPricer import *
 
-S0 = 431.3399963378906
-r = 0.02672004830382966
-q = 0.02
+data_folder = "test-0718/"
 
-dynamics = ["GBM", "HES", "MER"]
-n_inputs = {"GBM": 5, "HES": 9, "MER": 8}
-put_call = ["put", "call"]
-model_dim = [120,120,120,120,120,1]
-model_file = "model_500ep|120,120,120,120,120,1|l,l,l,l,l,i.pt"
-option_data = "../bkts_data/option_chain_SPY_2021-07-16.csv"
+args = {
+    "S0": 431.3399963378906,
+    "r": 0.02672004830382966,
+    "q": 0.02,
+    "M_range": [0.8, 1.2],
+    "dynamics": ["GBM", "HES", "MER"],
+    "n_inputs": {"GBM": 5, "HES": 9, "MER": 8},
+    "put_call": ["put", "call"],
+    "model_dim": [120,120,120,120,120,1],
+    "model_file": "model_500ep|120,120,120,120,120,1|l,l,l,l,l,i.pt",
+    "option_data": "../bkts_data/option_chain_SPY_2021-07-16.csv",
+    "init_vals": {"GBM": [0.2], "HES": [0.2,1,0.04,0.5,-0.5], "MER": [0.2,1,0.1,0.2]}
+}
 
-def main():
-    dyn = "HES"
-    pc = "put"
-    data = pd.read_csv(option_data)
+output_cols = [
+    "Contract Name",
+    "Type",
+    "Put/Call",
+    "Strike",
+    "Maturity (Year)",
+    "Market Price",
+    "DNN Price (Nelder-Mead)",
+    "DNN Price (Powell)",
+    "DNN Price (COBYLA)",
+    "Implied Vol",
+    "Delta",
+    "Gamma",
+    "Vega",
+    "Rho",
+    "Theta"
+]
+
+def calibrate_dnn(dyn, pc):
+    print("calibrating %s %s" % (dyn, pc))
+    data = pd.read_csv(args["option_data"])
     data = data[data["Put/Call"]==pc.capitalize()]
     K = data["Strike"].values
+    M = args["S0"]/K
+    data = data[(M >= args["M_range"][0]) & (M <= args["M_range"][1])]
+    K = data["Strike"].values
     T = data["Maturity (Year)"].values
     V = data["Market Price"].values
     n = len(data)
-    model_path = "result/%s trained networks/%s/%s" % (dyn, pc, model_file)
-    model = MLP(n_inputs[dyn], model_dim)
+    print("mkt data to calibrate:")
+    print(data.head(10))
+    MTtensor = torch.t(Tensor([args["S0"]/K,T]))
+    model_path = "result/%s trained networks/%s/%s" % (dyn, pc, args["model_file"])
+    model = MLP(args["n_inputs"][dyn], args["model_dim"])
     model.load(model_path)
-    def dnn_pricer(sig,kappa,theta,zeta,rho):
-        inputs = torch.cat((torch.t(Tensor([S0/K,T])), Tensor([[r,q,sig,kappa,theta,zeta,rho]]*n)), dim=1)
+    def dnn_pricer(x):
+        inputs = torch.cat((MTtensor, Tensor([np.concatenate(([args["r"],args["q"]],x))]*n)), dim=1)
         return K * model(inputs).detach().numpy().flatten()
     def objective(x):
-        return np.sqrt(mean_squared_error(dnn_pricer(*x), V))
-    init_vals = [0.2,1,0.04,0.5,-0.5]
-    print("initial objective", objective(init_vals))
-    # bounds = [(0,2), (0,100), (0,4), (0,2), (-1,1)]
-    # opt_res = spo.minimize(objective, x0=init_vals, bounds=bounds, method="Newton-CG")
-    for m in ["Nelder-Mead","Powell","COBYLA"]:
-        print("using %s" % m)
+        return np.sqrt(mean_squared_error(dnn_pricer(x), V))
+    init_vals = args["init_vals"][dyn]
+    print("initial objective:", objective(init_vals))
+    opt_log = dict()
+    for m in ["Nelder-Mead", "Powell", "COBYLA"]:
+        print("using method %s" % m)
         opt_res = spo.minimize(objective, x0=init_vals, method=m)
         params = opt_res.x
-        print(params, objective(params))
-        print(dnn_pricer(*params), V)
+        data["DNN Price (%s)" % m] = dnn_pricer(params)
+        opt_log[m] = {
+            "opt_res": opt_res,
+            "params": params
+        }
+        print("params:", params)
+        print("objective:", objective(params))
+    data = data[output_cols]
+    data.to_csv(data_folder + "dnn_%s_%s.csv" % (dyn, pc), index=False)
+    print()
 
-    # for dyn in dynamics:
-    #     for pc in put_call:
-    #         print("calibrating %s %s" % (dyn, pc))
-    #         model_path = "result/%s trained networks/%s/%s" % (dyn, pc, model_file)
-    #         model = MLP(n_inputs[dyn], model_dim)
-    #         model.load(model_path)
-    #         pass
+def main():
+    for dyn in args["dynamics"]:
+        for pc in args["put_call"]:
+            calibrate_dnn(dyn, pc)
 
 if __name__=="__main__":
     main()

dnn_pricing/result/mkt calibration/dnnMktCalibrate.out

@@ -0,0 +1,77 @@
+Gbm put calibration under different minimization methods
+
+initial objective: 6.868051664134034
+using method Nelder-Mead
+params: [0.21546875]
+objective: 6.731850438345304
+using method Powell
+params: [0.2154891]
+objective: 6.731850399083168
+using method COBYLA
+params: [0.21558086]
+objective: 6.731853992997851
+
+Gbm call calibration under different minimization methods
+
+initial objective: 5.072589837756422
+using method Nelder-Mead
+params: [0.16367188]
+objective: 3.8627662241380367
+using method Powell
+params: [0.16366962]
+objective: 3.862765750379424
+using method COBYLA
+params: [0.16356719]
+objective: 3.8627765314537537
+
+Merton put calibration under different minimization methods
+
+initial objective: 10.36094942879816
+using method Nelder-Mead
+params: [0.17151546 1.12247462 0.10837124 0.05494519]
+objective: 6.4849991226099135
+using method Powell
+params: [ 0.05557684  0.64628819 -0.03051313  0.30891579]
+objective: 5.629234594537098
+using method COBYLA
+params: [ 0.05326636  0.65409727 -0.03171914  0.30828395]
+objective: 5.629511560710035
+
+Merton call calibration under different minimization methods
+
+initial objective: 14.128172520193203
+using method Nelder-Mead
+params: [ 0.08454932  0.68137934 -0.14803701  0.13071758]
+objective: 3.342730860007632
+using method Powell
+params: [ 0.06462324  0.33818109 -0.21205753  0.2500078 ]
+objective: 3.5026590186496485
+using method COBYLA
+params: [ 0.07838252  0.90503712 -0.1276526   0.11451638]
+objective: 3.3674879165319225
+
+Heston put calibration under different minimization methods
+
+initial objective: 7.527090924736272
+using method Nelder-Mead
+params: [ 0.14294561  1.84320996  0.07317877 -0.02692959 -0.40631107]
+objective: 6.2212245738875955
+using method Powell
+params: [ 0.04538429  2.27430948  0.08126959  0.13789064 -0.91094551]
+objective: 6.128816126889603
+using method COBYLA
+params: [ 0.02980197  2.18469218  0.08623377  0.26658284 -0.44829848]
+objective: 6.213661333153228
+
+Heston call calibration under different minimization methods
+
+initial objective: 4.013464521974254
+using method Nelder-Mead
+params: [ 0.14836916  0.66304691  0.0680975   0.48761702 -0.65135713]
+objective: 3.313023758828946
+using method Powell
+params: [ 0.15182468  1.12138606  0.04178745  0.20569562 -1.18580852]
+objective: 3.209692306412288
+using method COBYLA
+params: [ 0.14916193  0.83663081  0.06301192  0.58304154 -0.54309511]
+objective: 3.328468441305359