Quick script to backtest Numerai's Crypto Meta Model

numerai_crypto_mm_tb10.py

import yfinance as yf
from numerapi import CryptoAPI
import pandas as pd
import numpy as np
import warnings
import datetime

api = CryptoAPI()
api.download_dataset(
	"crypto/v1.0/historical_meta_models.csv",
	"historical_meta_models.csv"
)

#load historical MM
mm = pd.read_csv('historical_meta_models.csv')
mm['date'] = pd.to_datetime( mm['date'] )

# get historical daily data for all tickers available in historical MM

all_data = []
for symbol in mm['symbol'].unique():
    print( symbol )
    data = yf.Ticker(f"{symbol}-USD")
    # first_date = datetime.datetime.utcfromtimestamp(data.info['firstTradeDateEpochUtc']).strftime('%Y-%m-%d')
    try:
        hist = data.history(
            # start=first_date, 
            start='2024-01-01',
            end=datetime.date.today().strftime('%Y-%m-%d'), 
            interval="1d"
        )
        hist['symbol'] = symbol
        if hist.iloc[-1].name.strftime('%Y-%m-%d') == ( datetime.date.today() - datetime.timedelta(days=1) ).strftime('%Y-%m-%d'):
            all_data.append( hist )
    except:
        continue

df = pd.concat( all_data )
df['date'] = pd.to_datetime( df.index.strftime('%Y-%m-%d') )
df.set_index('date')

# only keep weekday prices
df = df.loc[ df.index.weekday <= 4 ]

# compute future 20 day log returns
lag_days = 20
df['lr'] = df.groupby('symbol')['Close'].transform(
    lambda y: np.log(y).diff(lag_days).shift(-lag_days)
)

#join metamodel predictions with log returns and initialize portfolio weight vector

x = df.set_index(['symbol','date']).join( mm.set_index(['symbol','date']), how='outer' )
x = x[['lr','meta_model']].dropna()
x = x.loc[ x['lr'].abs() <= 1000 ]
x['w'] = 0

# Sort by date and meta_model, then assign weights based on the top and bottom ranks within each date group
x = x.sort_values(by=['date', 'meta_model'], ascending=False)

def assign_top_bottom_w(group):
    n = len(group)
    # Initialize all weights to 0
    group['w'] = 0
    # Assign 1 to the top 10 ranks and -1 to the bottom 10 ranks
    group.iloc[:10, group.columns.get_loc('w')] = 1
    group.iloc[-10:, group.columns.get_loc('w')] = -1
    # normalize portfolio weights
    group['w'] = (1/lag_days) * group['w'] / np.sum( np.abs( group['w'] ) )
    # print['w']
    return group

# Apply the function by date group
x = x.groupby('date', group_keys=False).apply(assign_top_bottom_w)

# view portoflio performance
x.groupby('date').apply( lambda y: y['lr'] @ y['w'] ).cumsum().plot(
    title='Numerai CRYPTO MM Long 10 Short 10 Equal Weight Dollar Neutral\n Cumulative Log Return @ 1X Leverage\n Only Resolved Rounds'
)