Algorithmic Trading: Signals, Backtesting, and What Quants Actually Do

The Promise and the Reality

Algorithmic trading — using computers to make trading decisions — accounts for a huge share of daily market volume. Some estimates put it at over 70% in major equity markets. The appeal is obvious: systematic strategies can process information faster, avoid emotional biases, and operate at scales impossible for human traders.

The reality is more nuanced. Building a profitable trading system is hard. Most ideas that look good on paper fail in practice. But understanding how it works — even at a high level — is essential for anyone in quant finance.

---

What Is a Trading Signal?

A signal (or alpha) is a quantitative indicator that predicts future returns. It could be:

• Momentum: stocks that went up recently tend to keep going up
• Mean reversion: stocks that deviated from their average tend to revert
• Value: stocks that look cheap on fundamental metrics tend to outperform
• Carry: assets with higher yields tend to outperform
• Sentiment: natural language processing of news, filings, social media

Each signal has a measurable information coefficient (IC) — the correlation between the signal and subsequent returns. An IC of 0.02 might sound tiny, but applied across hundreds of stocks, it can be meaningful.

---

From Signal to Strategy

A raw signal is not a strategy. You need:

1. Portfolio construction: convert signals into positions, accounting for risk, correlation, and constraints
2. Execution: how to actually trade without moving the market against you
3. Risk management: position sizing, stop losses, exposure limits
4. Transaction costs: commissions, slippage, market impact

A signal with a 5% annual return but 3% in transaction costs is not profitable. The gap between theoretical and realised returns is often where strategies live or die.

---

Backtesting: The Necessary Illusion

Backtesting applies a strategy to historical data to see how it would have performed. It is essential for development but riddled with traps:

The Dangers

Look-ahead bias: accidentally using information that was not available at the time. If your signal includes next quarter's earnings, you have cheated.

Survivorship bias: only including stocks that still exist today. Dead companies disappear from databases, and they were usually the losers. Ignoring them makes everything look better.

Overfitting: tuning parameters until the backtest looks perfect — and then watching the strategy fail in live trading. The more parameters you tune, the more you are fitting noise rather than signal.

Data snooping: testing hundreds of strategies and selecting the best. By pure chance, some will look amazing. This is the multiple testing problem from statistics — and it is the single biggest cause of failed trading strategies.

The Antidotes

• Use out-of-sample testing (train on one period, test on another)
• Apply realistic transaction costs
• Test across multiple markets and time periods
• Be suspicious of anything with a Sharpe ratio above 2 in backtests
• Use walk-forward analysis: repeatedly retrain and test on new data

---

Execution Algorithms

Once you have a strategy that generates trades, you need to execute them in the market. For small orders, this is trivial. For large orders, it is an optimisation problem: trade too fast and you move the market against yourself. Trade too slowly and the signal decays.

Common algorithms:

• TWAP (Time-Weighted Average Price): spread the order evenly over time
• VWAP (Volume-Weighted Average Price): trade proportionally to market volume
• Implementation Shortfall: minimise the gap between decision price and execution price
• Adaptive algorithms: adjust speed based on real-time market conditions

Execution quality can be the difference between a profitable strategy and a losing one.

---

Market Making vs Directional Trading

Not all algo strategies try to predict returns:

Market making profits from the bid-ask spread by continuously quoting prices. The challenge is managing inventory risk — if you accumulate too much of an asset, an adverse price move can wipe out your spread income. Firms like Citadel Securities, Optiver, and Flow Traders dominate this space.

Statistical arbitrage identifies pairs or baskets of related securities that have temporarily diverged and bets on convergence. It requires strong statistical skills and fast execution.

High-frequency trading (HFT) operates on microsecond timescales, exploiting tiny inefficiencies. The technology requirements are extreme — custom hardware, colocation, optimised network stacks. Rust and C++ are the languages of choice here.

---

Signal Decay

Most signals weaken over time as:

• More people discover them
• Markets adapt and price them in
• The economic regime changes

A signal that worked brilliantly in 2015 might be flat by 2025. This is why quant teams need to continuously research new signals — the alpha treadmill never stops.

---

Python for Strategy Research

Python is the dominant language for strategy research, thanks to Pandas and NumPy:

import pandas as pd
import numpy as np

# Simple momentum signal
prices = pd.read_csv("stock_prices.csv", parse_dates=["date"], index_col="date")
returns = prices.pct_change()

# Signal: 12-month return, skip most recent month
signal = prices.shift(21) / prices.shift(252) - 1  # ~1mo to ~12mo return

# Simple backtest: go long top quintile, short bottom quintile
ranks = signal.rank(axis=1, pct=True)
long_mask = ranks > 0.8
short_mask = ranks < 0.2

# Portfolio returns (equal weight within quintiles)
long_ret = (returns * long_mask).mean(axis=1)
short_ret = (returns * short_mask).mean(axis=1)
strategy_ret = long_ret - short_ret

---

Getting Into Algo Trading

Algorithmic trading sits at the intersection of statistics, programming, and market knowledge. It requires healthy scepticism (most ideas fail), rigorous testing, and a genuine understanding of market structure.

Quantt covers the full stack: the mathematical foundations you need, the programming skills to implement strategies, and the financial knowledge to understand what you are doing. All integrated, all interactive.

For further reading, Marcos López de Prado's blog is excellent on the pitfalls of backtesting, and the QuantConnect documentation provides a practical open-source platform for strategy development.