What Is Algorithmic Trading & Does It Really Work?
Algorithmic trading means using computer code to decide or place trades. That part is simple. The hard part is that the same two words cover a huge range of setups, and most of them should not be judged the same way.
A large bank uses it to break big orders into smaller pieces, so the market does not see them coming. A small fund uses it to run its whole process, from the trade idea to the final order, with almost no human touch. Both are called algorithmic trading. They are not the same business, and they do not carry the same risks.
You are reading this either because someone wants you to put money into an algorithm, or because you are thinking of building one yourself. The same evidence serves both: most of what gets sold, and most of what gets built at home, does not hold up once the hard questions start.
Three questions do most of the work: what does the code decide, how do live results compare with the research, and what happened in the last bad stretch.
What algorithmic trading means
Algorithmic trading is when software follows written rules to buy or sell, instead of a person making each call by hand. The rules can be narrow or wide. Some only handle the mechanics of placing an order. Others run the whole job, from picking the trade to sending it to the market.
The trouble is that one description fits setups with almost nothing in common. That has made the words useful cover: a manager running a mostly hand-picked strategy with good trading software can present himself the same way as one running a full model-driven engine. Once you set the sales pitch aside, most setups fall into three types.
| Type | What the code does | What that means for you |
|---|---|---|
| Execution-only | Times and routes orders to reduce market impact | The investment idea can still be made by hand; the code earns its keep through better fills, the prices orders end up getting. |
| Signal-driven | Produces the trade ideas or ranks opportunities | Here the research matters most: is there real out-of-sample evidence (results on data the model never saw), and do the returns survive going live? |
| Full-stack systematic | Runs signals, sizing, risk, and execution as one process | The best shot at a consistent process, and also the highest need for controls, infrastructure, and oversight. |
Three types, and three very different sets of questions. Start from the wrong one and every follow-up is off.
Has anyone made money with algo trading?
Yes, a small minority have, including people who built their own. Fewer than the marketing suggests: BarclayHedge’s Graveyard Database held 22,241 funds and programmes that closed or stopped reporting as of 2026, and the typical systematic programme lasts about five years. The survivors share habits you can check.
The rest fail sooner or later, and quietly. The record of closed programmes is the missing half of every success story, and it is the number the sales decks leave out.
Those numbers, from BarclayHedge’s Graveyard Database and the CTA survival literature, are the reason we start from doubt rather than trust.
How do I start algorithmic trading?
Starting is cheap: a retail platform, a data feed, and a strategy written down as rules you can test. Lasting is the expensive part. Before real money goes in, a beginner needs an honest backtest, realistic cost assumptions, and a plan for the day the system misbehaves.
The order matters. Write the rules first, in plain language, before any code: what to trade, when, how much, and when to stand aside. If the rules cannot be written down, they cannot be automated, and no platform fixes that.
How much money do you need to start algo trading?
Less for the account than most people think, and more for everything around it than most people budget. Retail brokers take small accounts, and free tiers exist for data and testing.
The real costs are decent data, the time to test honestly, and the discipline to keep positions small while the system proves itself. Thin budgets tend to skip exactly those, which is one reason the failure numbers above look the way they do.
Which platform to use is its own decision: algorithmic trading platforms covers what the credible ones share.
What to check before you trust your own build
Turn the questions from this page on your own screen. Did the backtest see the data it was built on, or was some of it held back? Would the results survive slower fills and wider spreads than the test assumed? What breaks first as the position grows?
A builder who cannot answer those sits in the same seat as a manager who cannot, with the added problem that nobody else is checking.
One newer route deserves a flag: yes, ChatGPT can write a trading algorithm, and thousands of people have asked it to. Code that runs is not the same thing as a strategy that survives testing, which makes the checks above matter even more for AI-written systems. We cover that ground under AI trading.
Why investors look at algorithmic trading
Investors look at algorithmic trading for three reasons: returns that move on their own path next to stocks and bonds, discipline that does not panic in a bad week, and reach no human can match. All three are real. None of them make an automated strategy safe by default.
The independent return is the main draw. It does not show up in every market or on demand, but it happens often enough, with enough live history behind it, to be worth a close look.
The discipline draw is quieter. A rule does not talk itself out of a trade halfway through a bad stretch, so a good strategy runs more cleanly when nobody is second-guessing it under pressure. That does not make the rule smart.
Reach is the third. Some strategies need to watch thousands of instruments at once, adjust every position against a risk budget, and react faster than any person can. Without automation those strategies cannot exist at all. Here the software is the cost of entry; the profit still has to come from somewhere else.
Automation does not fix a weak process. It runs it faster, and with more money behind it, which makes the damage worse. And most of what reaches the market sits at the weaker end.
- Most published algorithmic and quantitative strategies give back a large part of their tested returns when they run on fresh data, a result that holds across decades of peer-reviewed work in financial economics
- When Campbell Harvey and co-authors went through the published research, they counted at least 316 separate equity factors documented by 2016, and judged that most were probably false discoveries once you account for how many things had been tested
- A common rule among experienced research teams is to cut published backtest Sharpe ratios in half (the standard return-per-risk score) before treating them as evidence
- Winton, one of the more honest systematic managers on the public record, has said plainly that live trend-following (buying recent risers, selling recent fallers) tends to do worse than its backtest after launch, a more useful admission than most peers offer
“Automation does not turn a weak strategy into a strong one. It just lets a weak strategy lose money faster, with more of yours in it.”
Why do algorithmic trading strategies fail?
Four failures do most of the damage: trading costs the model never saw, size that changes the maths, controls that exist only on paper, and plumbing that breaks under stress. None of them are exotic. A strategy can look sharp in research and meet all four in its first live year.
The trading costs more than the model assumed
If a strategy trades often, works in thinner markets, or has to act when conditions turn rough, slippage opens a gap between the tested result and the real one. March 2020 was a plain reminder. Liquidity dried up, spreads jumped, and fill assumptions that looked fine on paper stopped holding.
When a manager calls their execution institutional quality, ask what that means in real fills, how they measure the gap between the model and live trading, and what the numbers say.
Getting bigger changes the maths
A strategy that works with a small amount often breaks under a large one. As the money grows, trades move the market more, more people crowd into the same ideas, and turnover gets expensive to absorb.
This is not the mistake of a few careless managers. It is the most common way a real early profit stream disappears before outside investors ever see it. The ones who handle it well planned for capacity early, before it started to hurt.
Either the controls are real or they are not
A rule-based process sounds safe. It is only as safe as the controls around it, and a system with weak controls simply makes its mistakes faster than a person would.
MiFID II Article 17, in force since January 3, 2018, spelled out what regulated firms across Europe must have: systems that hold up, risk limits set in advance, a tested way to switch trading off, and written change management.
The SEC’s Rule 15c3-5, with broad compliance required by July 14, 2011, set similar baselines for firms trading into U.S. markets. Both rules asked for what strong managers were already doing. They also showed how many were not.
FINRA’s 2015 Regulatory Notice 15-09 drew straight from examinations and settled cases: weak change control, too little testing before going live, poor monitoring afterwards. Ordinary process failures, run at scale.
A strong manager can tell you how model changes get approved, what is watched in real time, who can step in when live behaviour moves away from plan, and what the review looks like after something goes wrong. If those answers stay vague past the first real follow-up, the controls are probably thinner than the pitch suggests.
The plumbing is part of the investment
Bad data, dropped connections, thin market coverage, and weak error handling are not back-office trivia. They decide whether a strategy works at all. The Swiss franc shock of January 2015 made the point. The Swiss central bank dropped its currency cap in minutes, cross-currency trading seized up, and assumptions that had held for years became very expensive, very fast.
- Knight Capital, then one of the largest U.S. equity market makers, rolled out updated software on August 1, 2012 with a setup error that switched an old, dormant trading routine back on, with no working off-switch attached to it
- In the 45 minutes before the problem was contained, the system fired off millions of unintended trades across 154 stocks, building a position the firm had never meant to hold and could not afford
- The loss came to about $440 million in a single morning; the firm needed emergency funding and was bought by Getco.
- The SEC found Knight had failed to keep proper pre-trade risk controls on its automated system. The cause was basic deployment discipline and change management rather than an exotic technical failure.
What the better algorithmic trading managers do differently
The better algorithmic trading managers share four habits: they explain where the profit comes from in plain words, they name where the strategy fails without being pushed, they treat trading costs as part of the design, and they talk about capacity as what breaks first. The model itself is rarely what separates them.
The managers we have sat across from who hold up show these habits almost without exception.
The plain language is not dumbing down; they understand the strategy well enough to say it in a few sentences. A manager who needs ten minutes of build-up before reaching the point is no clearer in private.
Naming the failure spots unprompted tells you a great deal about whether the research was honest, or whether the strategy has only ever been shown in its best light.
The cost habit shows in the deck. The space between the modelled cost and the real cost is where most good-looking research stops looking good, and the strong ones closed that gap before anyone saw a slide.
On capacity, “We could handle another $200M” is a statement about appetite. “Our market-impact model starts to bend around here” is a statement about understanding: it names the size where costs start growing faster than the position.
In our experience, fewer than one in ten managers who arrive with a confident story can answer the oversight questions with the same ease. The gap usually shows inside the first half hour.
- A January 2025 SEC action against Two Sigma found the firm had left known access-control weaknesses in its investment models unfixed: $165 million repaid to clients and $90 million in penalties. The models were never the problem. Who could touch them was.
- Renaissance’s Medallion fund, the most successful systematic strategy on public record, has been closed to new outside money since 1993. The firm treats its own capacity as the thing to protect, which says more about size limits than any pitch deck
- Man AHL has run systematic programmes for more than three decades across regime changes. Continuity like that is the exception, and very few firms can show it
Questions worth asking before you trust a strategy
The questions below work across a table from a manager, and they work on your own build too.
About the strategy
- What does the code do? Place orders, generate signals, manage risk, or all of it?
- Which market conditions hurt it most?
- When did live performance last disappoint, and what happened?
About how it trades
- How are trading costs modelled, and how do real fills compare with those assumptions?
- What happened when liquidity got thin or prices moved fast?
- How is slippage measured in live trading, and who checks it?
About size
- What breaks first as the money grows?
- Where has capacity pressure shown up, and at what size?
- How realistic are the capacity numbers in the current documents?
About controls
- How are model changes approved, and who can say no?
- What is watched in real time, and by whom?
- What happened after the last real problem: a data failure, an odd fill, or unexpected live behaviour?
- How much rides on a few people, brokers, venues, or data providers?
| Common claim | What it often hides | What to check |
|---|---|---|
| Proprietary models | Complexity with no proof it works out of sample | Out-of-sample evidence, how model changes are handled, how many versions were tried before this one |
| Institutional-grade execution | Standard broker tools presented as an in-house capability | Slippage reporting, the benchmark used, how live fills compare with the research |
| Scalable capacity | Hopeful assumptions about fills at larger size | Turnover, how liquid the market is, market-impact modelling, a stated capacity limit |
| Risk management | Fixed limits instead of a real control system | What controls exist, who owns the off-switch, how incidents are reviewed, how limits are set |
| Low-correlation returns | A short track record or a hidden bet on one common driver | How it behaved under stress, what it quietly rises and falls with, what drove the losses. Correlations that look low in calm markets have a habit of rising when it matters most. |
Algorithmic trading and quant trading: how the terms differ
These words overlap, but they are not the same.
| Term | What it usually describes |
|---|---|
| Algorithmic trading | Using rules to make or place trades, from simple order tools to a fully systematic strategy |
| Quant trading | A more model-driven, data-heavy way of generating signals and building the portfolio |
| High-frequency trading | A narrow part of algorithmic trading built around speed and exchange mechanics |
| Systematic investing | A broader rules-based process, often reaching into portfolio construction and asset allocation |
Continue reading: Algorithmic trading strategies
Common questions about algorithmic trading
It means a computer follows written rules to decide or place trades, instead of a person handling each one. The rules can be as narrow as routing one order neatly, or as wide as running the whole process from idea to trade. Same name, very different systems underneath.
A strategy is written down as clear rules: what to trade, how much, when, and where. Software watches the market, checks those rules, and sends orders on its own when the conditions are met, with risk limits and monitoring sitting on top. How much runs without a person varies by firm.
Yes. Automated trading is legal in every major market and used every day by banks, funds, and exchanges. What regulators control is how it is run. Europe’s MiFID II Article 17 and the SEC’s Rule 15c3-5 require risk controls, testing, and a working off-switch from firms that use it.
High-frequency trading is one narrow part of algorithmic trading, built around very short holding times, servers placed inside the exchange’s own data centre, and speed. Most algorithmic strategies work over longer horizons and care more about signal quality and size discipline than speed. Treating the two as one leads to the wrong questions.
Ask three things: what does the code do, in one sentence; how does live performance compare with the research; and what happened in the last bad stretch. A manager who answers all three clearly, without dodging, is already ahead of most.
The main ones are hidden costs and false confidence. Real fills cost more than backtests assume, getting bigger changes the maths, and a coding or deployment error executes at machine speed. The deeper drawback: a system can look disciplined while automating a process that was never sound.
Because a backtest uses the same data the strategy was built on, assumes fills that live markets rarely give, and cannot model what happens when other strategies react to the same signals at the same time. Live trading corrects all of that at once, and the gap is usually wider than the research suggests.
Where we score real strategies
The questions on this page are the ones we score in The Review: twenty-five checks across risk, protection controls, execution, operations, and research discipline, weighed against the evidence the developer can produce. The scoring standard is written up in our review methodology. Most of what crosses our desk does not pass.