Algorithmic Trading

Overview

Build, backtest, and deploy algorithmic trading strategies. Cover market data ingestion, strategy development, backtesting with realistic assumptions, risk management, and live execution.

Instructions

Architecture

Market Data → Data Pipeline → Strategy Engine → Risk Manager → Order Executor → Exchange
     ↑                              ↓                                    ↓
     └──────────── Performance Monitor ←────── Position Tracker ←────────┘

Data pipeline

# data_fetcher.py — Fetch historical OHLCV data

import ccxt
import yfinance as yf

def fetch_crypto_ohlcv(symbol: str, timeframe: str, since: str) -> list:
    """Fetch candles from Binance. Returns [timestamp, O, H, L, C, V]."""
    exchange = ccxt.binance()
    ohlcv = exchange.fetch_ohlcv(symbol, timeframe,
                                  since=exchange.parse8601(since), limit=1000)
    return ohlcv

def fetch_stock_data(ticker: str, period: str = '2y', interval: str = '1d'):
    """Fetch stock OHLCV from Yahoo Finance."""
    return yf.download(ticker, period=period, interval=interval)

For real-time data, use WebSocket feeds (ccxt.pro):

async def stream_orderbook(symbol: str = 'BTC/USDT'):
    exchange = ccxtpro.binance()
    while True:
        ob = await exchange.watch_order_book(symbol)
        spread = (ob['asks'][0][0] - ob['bids'][0][0]) / ob['bids'][0][0] * 100
        print(f"{symbol} Spread: {spread:.4f}%")

Strategy development

Common types: Momentum/trend following (MAs, RSI, MACD), mean reversion (z-scores, Bollinger Bands), statistical arbitrage (pairs trading, cross-exchange), market making (spread capture).

# strategy.py — Dual Moving Average Crossover with RSI filter

import pandas as pd

def calculate_signals(df: pd.DataFrame, fast=20, slow=50, rsi_period=14):
    """Generate trading signals from OHLCV data."""
    df['sma_fast'] = df['close'].rolling(window=fast).mean()
    df['sma_slow'] = df['close'].rolling(window=slow).mean()

    delta = df['close'].diff()
    gain = delta.where(delta > 0, 0).rolling(window=rsi_period).mean()
    loss = (-delta.where(delta < 0, 0)).rolling(window=rsi_period).mean()
    df['rsi'] = 100 - (100 / (1 + gain / loss))

    df['signal'] = 0
    df.loc[(df['sma_fast'] > df['sma_slow']) &
           (df['sma_fast'].shift(1) <= df['sma_slow'].shift(1)) &
           (df['rsi'] < 70), 'signal'] = 1   # Buy
    df.loc[(df['sma_fast'] < df['sma_slow']) &
           (df['sma_fast'].shift(1) >= df['sma_slow'].shift(1)) &
           (df['rsi'] > 30), 'signal'] = -1  # Sell
    return df

Backtesting

A backtest must model real conditions — fees, slippage, and stop-losses:

# backtester.py — Vectorized backtester with realistic assumptions

class BacktestConfig:
    commission: float = 0.001     # 0.1% per trade
    slippage: float = 0.0005      # 0.05% estimated
    initial_capital: float = 10000
    position_size: float = 0.1    # 10% of portfolio per trade
    stop_loss: float = 0.02       # 2% stop-loss
    take_profit: float = 0.06     # 6% take-profit (3:1 R/R)

Key metrics:

RETURNS: Total Return, Annualized Return, Alpha (vs buy-and-hold)
RISK: Max Drawdown (<15%), Sharpe Ratio (>1.0 good, >2.0 excellent), Sortino, Calmar
TRADING: Win Rate (>45% for trend, >55% for mean reversion), Profit Factor (>1.5)

Risk management

Position sizing: Fixed fractional (X% per trade) or Kelly Criterion (f = (bp - q) / b, use half-Kelly for safety).

# risk_manager.py — Enforce risk rules before every order

RISK_RULES = {
    'max_position_pct': 0.10,       # Max 10% per position
    'max_portfolio_risk': 0.02,     # Max 2% risk per trade
    'max_daily_loss': 0.05,         # Stop after 5% daily loss
    'max_drawdown': 0.15,           # Stop after 15% drawdown
    'max_correlation': 0.7,         # No correlated positions
}

Live execution

# executor.py — Live orders with safety checks

def place_order(exchange, symbol, side, amount, order_type='limit', price=None):
    ticker = exchange.fetch_ticker(symbol)
    if order_type == 'market':
        spread = (ticker['ask'] - ticker['bid']) / ticker['bid'] * 100
        if spread > 0.5:
            raise ValueError(f"Spread too wide: {spread:.2f}%")
    if order_type == 'limit' and price:
        deviation = abs(price - ticker['last']) / ticker['last'] * 100
        if deviation > 2.0:
            raise ValueError(f"Limit price {deviation:.1f}% from market")
    return exchange.create_order(symbol, order_type, side, amount, price)

Always paper trade first: Alpaca (built-in), Binance Testnet, Interactive Brokers (TWS). Run 1 month or 100 trades minimum before live capital.

Examples

Build a crypto momentum strategy

Build a momentum trading strategy for BTC/USDT on Binance. Use EMA crossover (12/26) with volume confirmation and RSI filter. Backtest on 2 years of hourly data with realistic fees (0.1% taker), calculate Sharpe ratio and max drawdown, and compare against buy-and-hold. Include stop-loss at 2% and take-profit at 6%.

Create a pairs trading system

Build a statistical arbitrage system that trades correlated stock pairs. Use cointegration testing to find pairs from the S&P 500, implement a z-score mean reversion strategy, and backtest with transaction costs. Include the pair selection process, entry/exit rules, and risk management.

Set up a live trading bot with risk management

Deploy a live trading bot on Alpaca for US stocks. It should run a simple momentum strategy on a universe of 20 liquid ETFs, execute via limit orders, enforce position size limits (max 10% per holding), and stop trading if daily loss exceeds 2%. Include monitoring, logging, and alerting via Telegram.

Guidelines

• Always backtest with realistic fees, slippage, and stop-losses — unrealistic backtests are worthless
• Never skip paper trading — run for at least 1 month or 100 trades before deploying live capital
• Use half-Kelly criterion for position sizing to add a safety margin
• Implement hard daily loss limits and max drawdown circuit breakers
• Check spread width before market orders — never market-order into thin books
• Compare every strategy against buy-and-hold; if it underperforms, it's not worth the complexity
• Log every order, signal, and risk check for post-mortem analysis