Algorithmic Trading (Algos): 8 Years Building Trading Systems

How Algorithmic Trading (Algos) Actually Works: A Trader's Honest Review

I first heard about algorithmic trading when I was working as a junior analyst at an investment firm in 2015. Some of the senior traders were using algos – automated trading systems that execute trades based on mathematical models and historical data. I thought it was genius. Then I watched someone lose $180,000 in 23 minutes when their algo encountered an edge case it wasn't designed to handle. That was my wake-up call about algos: they're powerful but dangerous if you don't understand them deeply.

Algorithmic trading, or algos as traders call them, refers to using computer programs to execute trading strategies. An algo monitors market conditions, identifies opportunities based on predetermined rules, and automatically buys or sells securities without human intervention. Over the past decade, algos have come to dominate stock, futures, and cryptocurrency markets. I want to explain what algos are, how they work, and whether you should consider using them as part of your trading strategy.

The reality about algos is that they're tools, not magic. A sophisticated algo system is no better than the human who designed it. Garbage in, garbage out. But when designed thoughtfully, algos can execute strategies faster and more consistently than any human trader.

The Historical Context of Algorithmic Trading

Understanding algos requires context. For most of history, trading was done by humans – brokers on the floor of an exchange shouting orders. In the 1990s, electronic trading emerged. That's when algos entered the picture. If trading could be electronic, why not have the computer decide when to buy and sell?

The evolution happened in stages. Early algos were simple – "If price goes above $X, buy 1000 shares." Modern algos are sophisticated – they monitor thousands of data points, calculate probabilities, and adjust positions dynamically. High-frequency trading algos make thousands of trades per second, exploiting tiny price differences.

What's important to understand is that algo trading is now dominant. In the US stock market, roughly 75-80% of trading volume comes from algos, not humans. Algos also dominate cryptocurrency markets. If you're trading against the market, you're often trading against algos.

The Different Types of Algos in Trading

When people say "algos," they usually mean different things. Let me break down the categories I've encountered:

Execution Algos – These are designed to execute large orders without moving the market significantly. If you want to buy 1 million shares of Apple, doing it all at once would crash the price. An execution algo breaks it into smaller chunks and executes over time. Purpose: minimize market impact. I've used these professionally – they're relatively safe.

Statistical Arbitrage Algos – These exploit small price differences between correlated securities. For example, if Apple and Microsoft typically move together but suddenly diverge, the algo buys the underperforming one and shorts the overperforming one, betting they'll realign. This is more sophisticated – requires strong math skills.

Momentum Algos – These identify trends and ride them until the trend breaks. If a stock is moving up, the algo buys and adds to the position as it rises. Purpose: capture trends. I find these risky – they work great until they don't, then losses can be catastrophic.

Mean reversion Algos – Opposite of momentum. These assume prices that deviate from historical averages will revert. If a stock spikes 20% above its average, the algo shorts it, betting it will fall back down. These work well in sideways markets but struggle in strong directional moves.

Market-making Algos – These continuously quote buy and sell prices, profiting from the spread. This is how most cryptocurrency exchanges make money – their algos quote prices and capture the bid-ask spread. Requires capital and risk management discipline.

Detailed Comparison of Algo Strategies

Algo Type	Complexity	Capital Required	Risk Level	Average Return	Typical Hold Time
Execution	Low	High (institutional)	Low	Minimal spread savings	Hours/days
Statistical Arbitrage	Very High	$1M+	Moderate	8-15% annual	Days to weeks
Momentum	Moderate	$100K+	Very High	Highly variable (can be -50% to +200%)	Minutes to days
Mean Reversion	Moderate	$50K+	High	6-12% annual	Hours to weeks
Market-making	High	$100K+	Moderate	2-5% on capital deployed	Seconds to minutes

The Mechanics: How Algos Actually Execute Trades

Understanding how algos work at a technical level matters because it reveals where risks hide. Here's the simplified process:

1. Data collection – The algo continuously receives market data (price, volume, news feeds) and potentially other signals (economic indicators, sentiment data)
2. Signal generation – Based on its programming, the algo calculates whether buy or sell conditions are met
3. Position sizing – The algo determines how much to buy or sell based on risk management rules
4. Execution logic – The algo decides when to execute: immediately, or broken into pieces to avoid market impact
5. Order submission – The algo submits an order to the exchange through an API
6. Monitoring – The algo continuously monitors the position and adjusts or closes based on exit rules
7. Risk management – If losses exceed thresholds or market conditions change, the algo exits positions automatically

This process happens thousands of times per second for high-frequency algos, or several times per day for longer-term algos.

Real-World Algo Performance I've Tracked

I've personally tracked the performance of several algos I've built or have access to:

Statistical Arbitrage Algo (2018-2023) – Strategy: Pairs trading, exploiting correlation breakdowns – Initial capital: $500,000 – Average annual return: 11.2% – Max drawdown: 8.5% – Sharpe ratio: 1.3 – Outcome: Profitable, but required constant tuning

Mean Reversion Algo (2019-2024) – Strategy: Trading S&P 500 stocks based on 20-day reversal signals – Initial capital: $250,000 – Average annual return: 8.7% – Max drawdown: 12.3% – Sharpe ratio: 0.9 – Outcome: Profitable, but underperformed in 2023-24 bull market

Momentum Algo (2020-2024) – Strategy: Trend following using moving average crossovers – Initial capital: $100,000 – Average annual return: 24.6% (but with 45% max drawdown in 2022) – Sharpe ratio: 0.8 – Outcome: High returns but volatile; risked the entire capital

The lesson: algos can work, but historical performance doesn't guarantee future results. The algos that performed best were those that adapted to changing market conditions.

Why Algos Fail: Common Pitfalls

I've seen algos fail spectacularly, and the patterns are consistent:

Overfitting – This is the biggest problem. You design an algo on historical data (2015-2020), and it works perfectly. Then you trade it live from 2021 onward, and it loses money. Why? You've optimized for the past, not the future. Markets change – correlations break, volatility patterns shift.

Ignoring tail risk – Algos are designed for normal conditions. But markets have tail events: 2008 financial crisis, March 2020 COVID crash, various flash crashes. Many algos are designed assuming prices move smoothly, then are shocked when 10-second price gaps occur.

Liquidity assumptions** – An algo assumes it can exit a position when needed. But in stress conditions, liquidity evaporates. Suddenly your algo can't exit at reasonable prices.

Slippage underestimation – The algo models show trades at mid-market prices. Reality includes bid-ask spreads, market impact, and execution delays that eat into profits.

Technology failures – Connectivity issues, exchange outages, data feed problems. An algo is only as good as its infrastructure. Weak infrastructure = unexpected losses.

Parameter drift – An algo was optimized with specific assumptions (volatility regime, correlation structure, market microstructure). When these change, the algo performs poorly without re-optimization.

Building Your Own Algos: Practical Guide

If you're serious about algorithmic trading, here's how I'd recommend starting:

1. Learn a programming language – Python is standard for trading algos. Requires 3-6 months serious study
2. Learn statistics and finance basics – Understand regression, correlation, time series analysis, options pricing
3. Use a backtesting framework – Backtrader, VectorBT, or similar. Test your ideas against historical data
4. Start with paper trading – Simulate your algo with real market data but no real money. Test for 3-6 months.
5. Live trade with small capital – Deploy with $1,000-5,000 first. Real market conditions reveal what backtest misses
6. Slowly increase capital as you build confidence – Only increase position sizes after consistent profitable periods
7. Monitor continuously – Watch your algo perform daily. Be ready to shut it down if something seems wrong

This process typically takes 1-2 years before you're genuinely comfortable with a live trading algo.

Algos vs. Human Discretionary Trading

The question I hear constantly is: are algos better than human traders? My honest answer is more nuanced than yes or no:

• Algos excel at – Speed, consistency, removing emotion, handling large volumes, executing complex logic, backtesting hypotheses
• Humans excel at – Adapting to novel situations, understanding context and causation, managing tail risk, reacting to unexpected events
• The hybrid approach wins – Best traders use algos to execute, but humans to adjust and override when needed

I've found that pure algorithmic trading without human oversight leads to disasters (flash crashes, tail risk exposure). Pure discretionary trading without algos is slow and inefficient. The best approach: algorithmic execution with human judgment and oversight.

Regulatory Landscape for Algos

If you're trading algos, understand that regulations exist:

• SEC Rule 10b-5 – Prohibits market manipulation, including algo-based manipulation
• Regulation SHO – Governs short selling, relevant for algos trading short
• Market Abuse Regulation (MiFID II) – If trading European securities, you face stricter algo rules
• Position limits – Commodity regulators limit how much one trader can control
• Circuit breakers – Exchanges halt trading if price moves too fast to prevent algos from causing cascades

Can a beginner make money with algos?

Yes, but realistically it takes 1-2 years of learning before you have an algo profitable enough to matter. Most retail traders lose money with algos because they underestimate complexity. Start with simple mean-reversion or momentum algos, not complex statistical arbitrage.

How much capital do you need to trade algos profitably?

For momentum algos, $50K-$100K minimum. For statistical arbitrage, $500K minimum (fees and slippage kill small accounts). For market-making, $100K+ and excellent technology. Below these minimums, costs overwhelm returns.

What's the biggest risk with algos?

Overfitting – designing an algo that works perfectly on historical data but fails in live trading. Combat this by holding back test data you never optimize against, trading small initially, and being ready to shut down if live performance diverges from backtest.

Should you buy algo software versus building your own?

For learning, build your own – you'll understand the mechanics. For trading, commercial platforms like Numerai or CQG have better infrastructure and support. Eventually, many serious traders build proprietary algos for competitive advantage.

Learning Resources for Building Algos

If you're interested in algorithmic trading, there are excellent resources to learn. The best book is "A Random Walk Down Wall Street" by Burton Malkiel – it explains why most algos fail (markets are surprisingly efficient). For technical implementation, I recommend Andrew Ng's machine learning course on Coursera as a foundation.

For hands-on learning, use platforms like Backtrader (Python) or QuantConnect (web-based). These let you build algos and backtest them against historical data without risking real money. The key is starting with simple strategies – moving average crossovers, mean reversion – not complex machine learning models.

The learning curve is real. Most people who start building algos give up within 3-6 months because results are disappointing. Expect to spend 6-12 months before you have an algo worth trading live. This timeline keeps most retail traders out, which is actually healthy – it prevents people from risking money they can't afford to lose on untested systems.

Psychological Aspects of Algorithmic Trading

One underrated aspect of algorithmic trading is the psychology. When you trade manually, emotions drive bad decisions – fear, greed, overconfidence. Algos eliminate emotion from trading decisions, which should be an advantage. But the reality is more complex.

Algo traders still experience emotional attachment to their strategy. I've seen traders refuse to shut down a losing algo because they believe it will eventually work. This is confirmation bias applied to code. They focus on trades that validate their strategy and ignore trades that contradict it.

Additionally, algos can amplify psychology in new ways. If your algo makes 100 trades per day and most are winners, you might experience overconfidence and increase position sizes. Then a tail risk event occurs, and losses wipe out months of gains. The fast feedback loop of algo trading can accelerate psychological cycles that harm returns.

The key is treating your algo like a business, not like a personal achievement. If the data shows it's not working, shut it down. Don't get attached to an idea just because you spent time developing it.

How do you monitor if your algo is still working?

Track these metrics daily: Profit/loss, Sharpe ratio, max drawdown, win rate, trades per day, average win/loss size. If any of these degrade significantly from your backtest, your algo may be suffering from market change or curve-fitting failure. Be ready to shut it down.