Introduce fees into AgentMarket itself (not just OmenAgentMarket) (#…

…506)

.gitignore

@@ -162,3 +162,5 @@ cython_debug/
 
 tests_files/*
 !tests_files/.gitkeep
+
+bet_strategy_benchmark*

examples/monitor/match_bets_with_langfuse_traces.py

@@ -1,3 +1,4 @@
+from pathlib import Path
 from typing import Any
 
 import pandas as pd
@@ -89,6 +90,9 @@ def get_outcome_for_trace(
 
 
 if __name__ == "__main__":
+    output_directory = Path("bet_strategy_benchmark")
+    output_directory.mkdir(parents=True, exist_ok=True)
+
     # Get the private keys for the agents from GCP Secret Manager
     agent_gcp_secret_map = {
         "DeployablePredictionProphetGPT4TurboFinalAgent": "pma-prophetgpt4turbo-final",
@@ -240,7 +244,8 @@ def get_outcome_for_trace(
 
             details.sort(key=lambda x: x["sim_profit"], reverse=True)
             pd.DataFrame.from_records(details).to_csv(
-                f"{agent_name} - {strategy} - all bets.csv", index=False
+                output_directory / f"{agent_name} - {strategy} - all bets.csv",
+                index=False,
             )
 
             sum_squared_errors = 0.0
@@ -297,7 +302,9 @@ def get_outcome_for_trace(
             + simulations_df.to_markdown(index=False)
         )
         # export details per agent
-        pd.DataFrame.from_records(details).to_csv(f"{agent_name}_details.csv")
+        pd.DataFrame.from_records(details).to_csv(
+            output_directory / f"{agent_name}_details.csv"
+        )
 
     print(f"Correlation between p_yes mse and total profit:")
     for strategy_name, mse_profit in strat_mse_profits.items():
@@ -306,5 +313,7 @@ def get_outcome_for_trace(
         correlation = pd.Series(mse).corr(pd.Series(profit))
         print(f"{strategy_name}: {correlation=}")
 
-    with open("match_bets_with_langfuse_traces_overall.md", "w") as overall_f:
+    with open(
+        output_directory / "match_bets_with_langfuse_traces_overall.md", "w"
+    ) as overall_f:
         overall_f.write(overall_md)

prediction_market_agent_tooling/deploy/betting_strategy.py

@@ -4,7 +4,7 @@
 
 from prediction_market_agent_tooling.gtypes import xDai
 from prediction_market_agent_tooling.loggers import logger
-from prediction_market_agent_tooling.markets.agent_market import AgentMarket
+from prediction_market_agent_tooling.markets.agent_market import AgentMarket, MarketFees
 from prediction_market_agent_tooling.markets.data_models import (
     Currency,
     Position,
@@ -178,7 +178,7 @@ def calculate_trades(
         if self.max_price_impact:
             # Adjust amount
             max_price_impact_bet_amount = self.calculate_bet_amount_for_price_impact(
-                market, kelly_bet, 0
+                market, kelly_bet
             )
 
             # We just don't want Kelly size to extrapolate price_impact - hence we take the min.
@@ -196,15 +196,20 @@ def calculate_trades(
         return trades
 
     def calculate_price_impact_for_bet_amount(
-        self, buy_direction: bool, bet_amount: float, yes: float, no: float, fee: float
+        self,
+        buy_direction: bool,
+        bet_amount: float,
+        yes: float,
+        no: float,
+        fees: MarketFees,
     ) -> float:
         total_outcome_tokens = yes + no
         expected_price = (
             no / total_outcome_tokens if buy_direction else yes / total_outcome_tokens
         )
 
         tokens_to_buy = get_buy_outcome_token_amount(
-            bet_amount, buy_direction, yes, no, fee
+            bet_amount, buy_direction, yes, no, fees
         )
 
         actual_price = bet_amount / tokens_to_buy
@@ -216,7 +221,6 @@ def calculate_bet_amount_for_price_impact(
         self,
         market: AgentMarket,
         kelly_bet: SimpleBet,
-        fee: float,
     ) -> float:
         def calculate_price_impact_deviation_from_target_price_impact(
             bet_amount: xDai,
@@ -226,7 +230,7 @@ def calculate_price_impact_deviation_from_target_price_impact(
                 bet_amount,
                 yes_outcome_pool_size,
                 no_outcome_pool_size,
-                fee,
+                MarketFees.get_zero_fees(),  # TODO: Use market.fees
             )
             # We return abs for the algorithm to converge to 0 instead of the min (and possibly negative) value.
 

prediction_market_agent_tooling/jobs/omen/omen_jobs.py

@@ -75,7 +75,7 @@ def from_omen_agent_market(market: OmenAgentMarket) -> "OmenJobAgentMarket":
             market_maker_contract_address_checksummed=market.market_maker_contract_address_checksummed,
             condition=market.condition,
             finalized_time=market.finalized_time,
-            fee=market.fee,
+            fees=market.fees,
         )
 
 

prediction_market_agent_tooling/markets/agent_market.py

@@ -2,7 +2,7 @@
 from enum import Enum
 
 from eth_typing import ChecksumAddress
-from pydantic import BaseModel, field_validator
+from pydantic import BaseModel, field_validator, model_validator
 from pydantic_core.core_schema import FieldValidationInfo
 
 from prediction_market_agent_tooling.config import APIKeys
@@ -16,6 +16,7 @@
     ResolvedBet,
     TokenAmount,
 )
+from prediction_market_agent_tooling.markets.market_fees import MarketFees
 from prediction_market_agent_tooling.tools.utils import (
     DatetimeUTC,
     check_not_none,
@@ -60,6 +61,7 @@ class AgentMarket(BaseModel):
     current_p_yes: Probability
     url: str
     volume: float | None  # Should be in currency of `currency` above.
+    fees: MarketFees
 
     @field_validator("outcome_token_pool")
     def validate_outcome_token_pool(
@@ -77,6 +79,14 @@ def validate_outcome_token_pool(
                 )
         return outcome_token_pool
 
+    @model_validator(mode="before")
+    def handle_legacy_fee(cls, data: dict[str, t.Any]) -> dict[str, t.Any]:
+        # Backward compatibility for older `AgentMarket` without `fees`.
+        if "fees" not in data and "fee" in data:
+            data["fees"] = MarketFees(absolute=0.0, bet_proportion=data["fee"])
+            del data["fee"]
+        return data
+
     @property
     def current_p_no(self) -> Probability:
         return Probability(1 - self.current_p_yes)

prediction_market_agent_tooling/markets/manifold/manifold.py

@@ -6,6 +6,7 @@
 from prediction_market_agent_tooling.markets.agent_market import (
     AgentMarket,
     FilterBy,
+    MarketFees,
     SortBy,
 )
 from prediction_market_agent_tooling.markets.data_models import BetAmount, Currency
@@ -33,6 +34,13 @@ class ManifoldAgentMarket(AgentMarket):
     currency: t.ClassVar[Currency] = Currency.Mana
     base_url: t.ClassVar[str] = MANIFOLD_BASE_URL
 
+    # Manifold has additional fees than `platform_absolute`, but they don't expose them in the API before placing the bet, see https://docs.manifold.markets/api.
+    # So we just consider them as 0, which anyway is true for all markets I randomly checked on Manifold.
+    fees: MarketFees = MarketFees(
+        bet_proportion=0,
+        absolute=0.25,  # For doing trades via API.
+    )
+
     def get_last_trade_p_yes(self) -> Probability:
         """On Manifold, probablities aren't updated after the closure, so we can just use the current probability"""
         return self.current_p_yes

prediction_market_agent_tooling/markets/market_fees.py

@@ -0,0 +1,36 @@
+from pydantic import BaseModel, Field
+
+
+class MarketFees(BaseModel):
+    bet_proportion: float = Field(
+        ..., ge=0.0, lt=1.0
+    )  # proportion of the bet, from 0 to 1
+    absolute: float  # absolute value paid in the currency of the market
+
+    @staticmethod
+    def get_zero_fees(
+        bet_proportion: float = 0.0,
+        absolute: float = 0.0,
+    ) -> "MarketFees":
+        return MarketFees(
+            bet_proportion=bet_proportion,
+            absolute=absolute,
+        )
+
+    def total_fee_absolute_value(self, bet_amount: float) -> float:
+        """
+        Returns the total fee in absolute terms, including both proportional and fixed fees.
+        """
+        return self.bet_proportion * bet_amount + self.absolute
+
+    def total_fee_relative_value(self, bet_amount: float) -> float:
+        """
+        Returns the total fee relative to the bet amount, including both proportional and fixed fees.
+        """
+        if bet_amount == 0:
+            return 0.0
+        total_fee = self.total_fee_absolute_value(bet_amount)
+        return total_fee / bet_amount
+
+    def get_bet_size_after_fees(self, bet_amount: float) -> float:
+        return bet_amount - self.total_fee_absolute_value(bet_amount)

prediction_market_agent_tooling/markets/metaculus/metaculus.py

@@ -5,6 +5,7 @@
 from prediction_market_agent_tooling.markets.agent_market import (
     AgentMarket,
     FilterBy,
+    MarketFees,
     SortBy,
 )
 from prediction_market_agent_tooling.markets.metaculus.api import (
@@ -29,6 +30,7 @@ class MetaculusAgentMarket(AgentMarket):
     description: str | None = (
         None  # Metaculus markets don't have a description, so just default to None.
     )
+    fees: MarketFees = MarketFees.get_zero_fees()  # No fees on Metaculus.
 
     @staticmethod
     def from_data_model(model: MetaculusQuestion) -> "MetaculusAgentMarket":

prediction_market_agent_tooling/markets/omen/omen.py

@@ -22,6 +22,7 @@
 from prediction_market_agent_tooling.markets.agent_market import (
     AgentMarket,
     FilterBy,
+    MarketFees,
     SortBy,
 )
 from prediction_market_agent_tooling.markets.data_models import (
@@ -101,7 +102,6 @@ class OmenAgentMarket(AgentMarket):
     finalized_time: DatetimeUTC | None
     created_time: DatetimeUTC
     close_time: DatetimeUTC
-    fee: float  # proportion, from 0 to 1
 
     _binary_market_p_yes_history: list[Probability] | None = None
     description: str | None = (
@@ -240,7 +240,7 @@ def calculate_sell_amount_in_collateral(
             shares_to_sell=amount.amount,
             holdings=wei_to_xdai(pool_balance[self.get_index_set(sell_str)]),
             other_holdings=wei_to_xdai(pool_balance[self.get_index_set(other_str)]),
-            fee=self.fee,
+            fees=self.fees,
         )
         return xDai(collateral)
 
@@ -352,7 +352,12 @@ def from_data_model(model: OmenMarket) -> "OmenAgentMarket":
             url=model.url,
             volume=wei_to_xdai(model.collateralVolume),
             close_time=model.close_time,
-            fee=float(wei_to_xdai(model.fee)) if model.fee is not None else 0.0,
+            fees=MarketFees(
+                bet_proportion=(
+                    float(wei_to_xdai(model.fee)) if model.fee is not None else 0.0
+                ),
+                absolute=0,
+            ),
             outcome_token_pool={
                 model.outcomes[i]: wei_to_xdai(Wei(model.outcomeTokenAmounts[i]))
                 for i in range(len(model.outcomes))
@@ -598,7 +603,7 @@ def get_buy_token_amount(
             buy_direction=direction,
             yes_outcome_pool_size=outcome_token_pool[OMEN_TRUE_OUTCOME],
             no_outcome_pool_size=outcome_token_pool[OMEN_FALSE_OUTCOME],
-            fee=self.fee,
+            fees=self.fees,
         )
         return TokenAmount(amount=amount, currency=self.currency)
 
@@ -628,10 +633,10 @@ def get_new_p_yes(self, bet_amount: BetAmount, direction: bool) -> Probability:
         no_outcome_pool_size = outcome_token_pool[self.get_outcome_str_from_bool(False)]
 
         new_yes_outcome_pool_size = yes_outcome_pool_size + (
-            bet_amount.amount * (1 - self.fee)
+            self.fees.get_bet_size_after_fees(bet_amount.amount)
         )
         new_no_outcome_pool_size = no_outcome_pool_size + (
-            bet_amount.amount * (1 - self.fee)
+            self.fees.get_bet_size_after_fees(bet_amount.amount)
         )
 
         received_token_amount = self.get_buy_token_amount(bet_amount, direction).amount
@@ -1104,7 +1109,6 @@ def omen_remove_fund_market_tx(
     market_contract.removeFunding(api_keys=api_keys, remove_funding=shares, web3=web3)
 
     conditional_tokens = OmenConditionalTokenContract()
-    parent_collection_id = build_parent_collection_id()
     amount_per_index_set = get_conditional_tokens_balance_for_market(
         market, from_address, web3
     )
@@ -1116,7 +1120,6 @@ def omen_remove_fund_market_tx(
     result = conditional_tokens.mergePositions(
         api_keys=api_keys,
         collateral_token_address=market.collateral_token_contract_address_checksummed,
-        parent_collection_id=parent_collection_id,
         conditionId=market.condition.id,
         index_sets=market.condition.index_sets,
         amount=amount_to_merge,
@@ -1266,14 +1269,14 @@ def get_buy_outcome_token_amount(
     buy_direction: bool,
     yes_outcome_pool_size: float,
     no_outcome_pool_size: float,
-    fee: float,
+    fees: MarketFees,
 ) -> float:
     """
     Calculates the amount of outcome tokens received for a given investment
 
     Taken from https://github.com/gnosis/conditional-tokens-market-makers/blob/6814c0247c745680bb13298d4f0dd7f5b574d0db/contracts/FixedProductMarketMaker.sol#L264
     """
-    investment_amount_minus_fees = investment_amount * (1 - fee)
+    investment_amount_minus_fees = fees.get_bet_size_after_fees(investment_amount)
     buy_token_pool_balance = (
         yes_outcome_pool_size if buy_direction else no_outcome_pool_size
     )

prediction_market_agent_tooling/markets/omen/omen_contracts.py

@@ -163,10 +163,10 @@ def mergePositions(
         self,
         api_keys: APIKeys,
         collateral_token_address: ChecksumAddress,
-        parent_collection_id: HexStr,
         conditionId: HexBytes,
         index_sets: t.List[int],
         amount: Wei,
+        parent_collection_id: HexStr = build_parent_collection_id(),
         web3: Web3 | None = None,
     ) -> TxReceipt:
         return self.send(

prediction_market_agent_tooling/markets/polymarket/polymarket.py

@@ -3,6 +3,7 @@
 from prediction_market_agent_tooling.markets.agent_market import (
     AgentMarket,
     FilterBy,
+    MarketFees,
     SortBy,
 )
 from prediction_market_agent_tooling.markets.data_models import BetAmount, Currency
@@ -26,6 +27,12 @@ class PolymarketAgentMarket(AgentMarket):
     currency: t.ClassVar[Currency] = Currency.USDC
     base_url: t.ClassVar[str] = POLYMARKET_BASE_URL
 
+    # Based on https://docs.polymarket.com/#fees, there are currently no fees, except for transactions fees.
+    # However they do have `maker_fee_base_rate` and `taker_fee_base_rate`, but impossible to test out our implementation without them actually taking the fees.
+    # But then in the new subgraph API, they have `fee: BigInt! (Percentage fee of trades taken by market maker. A 2% fee is represented as 2*10^16)`.
+    # TODO: Check out the fees while integrating the subgraph API or if we implement placing of bets on Polymarket.
+    fees: MarketFees = MarketFees.get_zero_fees()
+
     @staticmethod
     def from_data_model(model: PolymarketMarketWithPrices) -> "PolymarketAgentMarket":
         return PolymarketAgentMarket(

prediction_market_agent_tooling/tools/betting_strategies/kelly_criterion.py

@@ -1,3 +1,4 @@
+from prediction_market_agent_tooling.markets.market_fees import MarketFees
 from prediction_market_agent_tooling.tools.betting_strategies.utils import SimpleBet
 
 
@@ -61,7 +62,7 @@ def get_kelly_bet_full(
     estimated_p_yes: float,
     confidence: float,
     max_bet: float,
-    fee: float = 0.0,  # proportion, 0 to 1
+    fees: MarketFees = MarketFees.get_zero_fees(),  # TODO: Remove default value.
 ) -> SimpleBet:
     """
     Calculate the optimal bet amount using the Kelly Criterion for a binary outcome market.
@@ -86,9 +87,14 @@ def get_kelly_bet_full(
     limitations under the License.
     ```
     """
+    fee = fees.bet_proportion
+    if fees.absolute > 0:
+        raise RuntimeError(
+            f"Kelly works only with bet-proportional fees, but the fees are {fees=}."
+        )
+
     check_is_valid_probability(estimated_p_yes)
     check_is_valid_probability(confidence)
-    check_is_valid_probability(fee)
 
     if max_bet == 0:
         return SimpleBet(direction=True, size=0)