1

banks.csv

@@ -0,0 +1,5 @@
+bank_id,CET1E,leverage,debt_sec,govbonds
+B01 25000 4.00 40000 20000
+B02 25000 6.00 40000 20000
+B03 25000 3.90 40000 20000
+B04 25000 3.10 40000 20000

markets.py

@@ -3,6 +3,7 @@
 import numpy as np
 
 from contracts import eps
+from behaviours import pay_off_liabilities
 
 # This represents a sale order in an asset market's orderbook.
 class Order:
@@ -19,7 +20,8 @@ def settle(self):
         # Sell the asset at the mid-point price
         value_sold = quantity_sold * (self.asset.price + old_price) / 2
         if value_sold >= eps:
-            self.asset.assetParty.add_cash(value_sold)
+            #self.asset.assetParty.add_cash(value_sold)
+            pay_off_liabilities(self.asset.assetParty, value_sold)
 
 
 # The key functions are clear_the_market() and compute_price_impact()

reinforcement_learning/MARL/NaiveA2C/ddpg_agent.py

@@ -179,12 +179,12 @@ def sample(self):
         """Randomly sample a batch of experiences from memory."""
         experiences = random.sample(self.memory, k=self.batch_size)
 
-        states = torch.from_numpy(np.vstack([e.state.astype(float) for e in experiences if e is not None])).float().to(device)
-        actions = torch.from_numpy(np.vstack([e.action.astype(float) for e in experiences if e is not None])).float().to(device)
-        rewards = torch.from_numpy(np.vstack([e.reward for e in experiences if e is not None])).float().to(device)
-        next_states = torch.from_numpy(np.vstack([e.next_state.astype(float) for e in experiences if e is not None])).float().to(
+        states = torch.from_numpy(np.asarray([e.state.astype(float) for e in experiences if e is not None], dtype='float64')).float().to(device)
+        actions = torch.from_numpy(np.asarray([e.action.astype(float) for e in experiences if e is not None], dtype='float64')).float().to(device)
+        rewards = torch.from_numpy(np.asarray([e.reward for e in experiences if e is not None], dtype='float64')).float().to(device)
+        next_states = torch.from_numpy(np.asarray([e.next_state.astype(float) for e in experiences if e is not None], dtype='float64')).float().to(
             device)
-        dones = torch.from_numpy(np.vstack([e.done for e in experiences if e is not None]).astype(np.uint8)).float().to(
+        dones = torch.from_numpy(np.asarray([e.done for e in experiences if e is not None], dtype='float64').astype(np.uint8)).float().to(
             device)
 
         return (states, actions, rewards, next_states, dones)

reinforcement_learning/MARL/NaiveA2C/model_train.py

@@ -17,21 +17,20 @@ def MA_obs_to_bank_obs(obs, bank):
 
 RLagent_dict = {}
 env = RLModelEnv()
-env.parameters.INITIAL_SHOCK = 0.01
-env.parameters.SIMULTANEOUS_FIRESALE = False
-bank_names = 'AT01 AT02 BE03 BE04 DK05 DK06 DK07 FI08 FR09 FR10 FR11 FR12 FR13 FR14 DE15 DE16 DE17 DE18 DE19 DE20 DE21 DE22 HU23 IE24 IE25 IT26 IT27 IT28 IT29 NL30 NL31 NL32 NL33 NO34 PL35 PL36 ES37 ES38 ES39 ES40 SE41 SE42 SE43 SE44 UK45 UK46 UK47 UK48'.split()
+
+bank_names = [f'B0{i}' for i in range(1, 5)]
 
 for idx, name in enumerate(bank_names):
     agent = Agent(state_size=3, action_size=2, random_seed=idx, name=name)
     RLagent_dict[name] = agent
 
 average_lifespans = []
-for episode in range(1000):
+for episode in range(10000):
 
     if episode == 0 or episode % 100 == 0:
         print(f'=========================================Episode {episode}===============================================')
     current_obs = env.reset()
-    play, max_play = 0, 15
+    play, max_play = 0, 5
     num_default = []
     while play < max_play:
         actions = {}

reinforcement_learning/gym_env.py

@@ -19,13 +19,8 @@ def choose_actions(self, action):
         # 0) If I'm insolvent, default.
         if self.is_insolvent():
             raise DefaultException(self, 'SOLVENCY')
-        balance = self.get_cash_()
         # 1. Pay off liabilities to delever
-        amountToDeLever = self.leverageConstraint.get_amount_to_delever()
-        if amountToDeLever > 0:
-            deLever = pay_off_liabilities(self, amountToDeLever)
-            balance -= deLever
-            amountToDeLever -= deLever
+        # Moved to Order.settle()
 
         # 2. Raise liquidity to delever later
         sellAssetActions = self.get_all_actions_of_type(SellAsset)
@@ -61,7 +56,7 @@ def reset(self):
         self.allAgents_dict = {}
         self.assetMarket = AssetMarket(self)
         obs = {}
-        with open('EBA_2018.csv', 'r') as data:
+        with open('/Users/mmw/Documents/GitHub/firesale_stresstest/banks.csv', 'r') as data:
             self.bank_balancesheets = data.read().strip().split('\n')[1:]
         for bs in self.bank_balancesheets:
             # This steps consist of loading balance sheet from data file
@@ -144,6 +139,8 @@ def step(self, action_dict):
 
     def stupid_action(bank):
         action = {}
+        CB_qty = bank.get_ledger().get_asset_valuation_of(Tradable, AssetType.CORPORATE_BONDS)
+        GB_qty = bank.get_ledger().get_asset_valuation_of(Tradable, AssetType.GOV_BONDS)
         action[AssetType.CORPORATE_BONDS], action[AssetType.GOV_BONDS] = 0.2 * abs(np.random.normal() - 0.5), 0.2 * np.random.normal() * abs(np.random.normal() - 0.5)
         return action
 
@@ -153,6 +150,12 @@ def stupid_action(bank):
         actions = {}
         play += 1
         for bank_name, bank in env.allAgents_dict.items():
+            CB_qty = bank.get_ledger().get_asset_valuation_of(Tradable, 1)
+            GB_qty = bank.get_ledger().get_asset_valuation_of(Tradable, 2)
+            equity = bank.get_ledger().get_equity_valuation()
+            lev_ratio_percent = bank.leverageConstraint.get_leverage() * 100
+            print(
+                f'Round {play}. Bank {bank_name}, CB: {int(CB_qty)}, GB: {int(GB_qty)}, EQUITY: {int(equity)}, LEV: {int(lev_ratio_percent)}%')
             actions[bank_name] = stupid_action(bank)  # this is where you use your RLAgents!
         obs, _, _, infos = env.step(actions)
         num_defaults.append(infos['NUM_DEFAULTS'])