Algorithmic trading uses computer programs to execute trades based on predefined rules and mathematical models. These algorithms can process vast amounts of data, identify patterns, and execute trades far faster than any human trader. Today, algorithms account for the majority of all trading volume on major exchanges.
Algorithms range from simple rule-based strategies (buy when the 50-day average crosses above the 200-day average) to extraordinarily complex machine learning models that analyze satellite imagery, natural language, and alternative data sources.
Here's the part beginners often miss: the computer isn't the edge. The computer just follows instructions without fear, boredom, or second-guessing. The edge, if there is one, lives in the rules themselves. A bad rule executed perfectly is still a bad rule.
Every trading algorithm does one of two jobs, and it helps to keep them separate in your head.
Alpha-generation algorithms decide what to buy or sell. Alpha is a finance word for return beyond what the market would have handed you anyway. These algorithms hunt for signals in prices, fundamentals, or alternative data.
Execution algorithms decide how to trade once the decision is made. Suppose a fund wants to buy a million shares. Dumping that order in at once would push the price up against them — that's called market impact. So execution algorithms slice big orders into small pieces. TWAP spreads pieces evenly over time. VWAP matches the market's natural volume rhythm, trading more when the market is busy. Implementation shortfall algorithms balance the cost of moving fast against the risk of the price drifting away while you wait.
Most of the algorithmic volume you hear about is execution, not alpha. It's plumbing — essential, but it isn't trying to predict anything.
Alpha strategies mostly fall into a few classic families.
Momentum strategies bet that recent winners keep winning for a while. Buy what's been going up, sell what's been going down, and rebalance regularly. Momentum has shown up across markets and decades in academic studies, but it suffers brutal reversals — momentum portfolios crashed hard in 2009 when beaten-down stocks snapped back.
Mean reversion strategies bet the opposite way over shorter windows: prices that stretch too far from their recent average tend to snap back. A stock that fell five percent today on no news often recovers part of that move. Mean reversion tends to win small amounts frequently, then lose big when a stretched price keeps stretching — because sometimes the drop is the news.
Statistical arbitrage (stat arb) trades relationships between securities rather than single names. The classic version is pairs trading: find two stocks that historically move together, and when they drift apart, buy the laggard and short the leader, betting the gap closes. Modern stat arb runs this idea across thousands of names at once, leaning on the law of large numbers.
High-frequency trading (HFT) is a subset of algorithmic trading that relies on extremely low latency and high throughput, executing thousands of trades per second to capture tiny price discrepancies. HFT firms invest heavily in co-located servers, specialized hardware, and network infrastructure to gain microsecond advantages.
Latency is just the delay between deciding to trade and the order reaching the exchange. Whether it matters depends entirely on how long your edge lasts. If your signal is a price discrepancy that vanishes in a millisecond, whoever gets there first takes the whole prize — it's a race, and second place earns nothing. That's why HFT firms pay for servers physically inside exchange buildings.
But if your signal plays out over weeks or months — like a value or momentum strategy — arriving a few seconds or even minutes later costs you almost nothing. This is genuinely good news for individual investors. You can't win a speed race against firms spending millions on fiber and hardware. You were never going to. Slower strategies don't require you to.
Most algorithmic strategies don't die dramatically. They fade.
One killer is overfitting: tuning rules so tightly to past data that they memorize old noise instead of learning a durable pattern. An overfit strategy looks brilliant in a backtest and ordinary — or worse — live.
Another is crowding. When a pattern becomes well known, many algorithms chase it, and their combined trading shrinks the very mispricing they were exploiting. Edges get competed away because that's what markets do.
A third is regime change. Rules calibrated in a calm, low-rate decade can break when volatility or policy shifts. The market that trained your model isn't guaranteed to be the market that grades it.
So treat any backtested edge — yours or anyone else's — as a hypothesis, not a promise. Expect live results to come in below the backtest. If a strategy only works when its assumptions hold perfectly, plan for the day they don't.
You don't need a server farm to think algorithmically. You need a rule you can write down, data to test it against, and honesty about the results.
Start with a rule simple enough to explain in one sentence. Backtest it across decades of history, not just the recent stretch where it happens to shine. Include trading costs, because a strategy that trades daily can see a paper edge disappear entirely into fees and slippage — the gap between the price you expected and the price you got.
Then try to break it. Test it on different years, different stocks, slightly different settings. A robust rule survives small changes; a fragile one collapses. This is the real work of algorithmic trading, and it's mostly the work of proving yourself wrong before the market does it for you — with your money.
Previous: Insider Trading Signals · Next: Quantitative Strategies · Financial glossary
← Back to all investing concepts