Algorithmic trading primer: Expert Advisors, backtests and overfitting

A trading rule has a condition and an action: when this is true, do that. A complete set also says how much to trade and when to get out. Written precisely enough for a computer to follow, it is an algorithm.

Automation removes hesitation, inconsistency and fatigue. It does not remove risk, and it adds some of its own: a program repeats a mistake as faithfully as it repeats anything else.

The term covers a wide range. Large institutions use algorithms mainly to carry out big orders in small pieces. What an individual meets is usually a rule-based strategy running on a retail platform.

Expert Advisor, or EA, is MetaTrader’s name for a program that is attached to a chart and allowed to trade. It is written in the platform’s own language, MQL5 in the case of MetaTrader 5. It can read prices and indicators, and send, change and close orders.

An EA runs inside the trading terminal. It works only while that terminal is running, connected to the server and permitted to trade automatically. Its code is called each time a new price arrives.

Two neighbours are often confused with it. An indicator calculates and draws but does not trade. A script runs once and stops.

An EA bought from a stranger comes with the seller’s own test results, and they are advertising. Nobody can tell from a results table how the rules were arrived at, and that, as the next two sections explain, is the thing that matters.

A backtest runs the rules over past prices to see what they would have done. It is useful for finding errors in the rules, for seeing how they behave, and for meeting their worst stretches before those cost anything. It misleads in well-known ways.

• The data. History may be incomplete, or built from bars with the movement inside each bar guessed. A rule that depends on what happened within a bar is then being tested on invention.
• Costs left out. Spread, commission, overnight financing and slippage are all real. A rule that makes many small trades can pass from profit to loss on the spread alone.
• Looking ahead. The test uses something that was not known at the time, such as a bar’s closing price to decide a trade made at its open.
• Fills that could not have happened: an exact price in the middle of a gap, or any size at the quoted price.
• Survivors only. A test run on instruments that exist today leaves out those that failed and were delisted.
• Too few trades. Thirty trades are an anecdote, however good they look.
• One kind of market. Rules tuned to a long quiet trend have never met anything else.

Optimisation tries many values for a rule’s settings and keeps those that did best. The danger is built in. Past prices are made of some pattern and a great deal of noise, and with enough attempts some combination will fit the noise by chance. It will look excellent on the past and mean nothing for the future. That is overfitting.

It shows itself in a few ways: a result that collapses when a setting is changed slightly; many settings and few trades; a line of results too smooth to believe.

The defences are habits and not proofs. Keep part of the history back, unseen, and test the chosen settings on it once: an out-of-sample test. Repeat that over successive stretches, choosing on one and testing on the next: walk-forward testing. Prefer few settings, and a broad region where results are acceptable to a single sharp peak. Then run the rules on a demonstration account, on prices that did not exist when they were written.

A rule that survives all of this has not been shown to work in future. It has only failed to be shown not to.

Latency is the time between the program deciding to trade and the order being executed. It is made up of the computer, the path across the network and the handling at the far end.

How much it matters depends on how long the opportunity lasts. To a rule that acts on four-hour bars, a tenth of a second is nothing. To a rule that tries to take a few points many times a day, it can be the whole result, and in that race an individual competes with firms whose machines sit beside the exchange’s own.

A virtual private server, or VPS, is a rented computer in a data centre that is always on and has a steady connection. The terminal runs there, so the EA keeps working when the computer at home is switched off or the line fails. A VPS deals with uptime and shortens the network path. It does nothing for the quality of the rules, it costs a monthly rent, and it needs looking after, since updates and restarts happen there too.

Real running has faults that a backtest lacks. Connections drop. Terminals restart, and a program must then work out what positions it has. An order is rejected or requoted, and the code must have an answer. A broker changes a spread, a minimum stop distance or a symbol’s name.

A program left unattended is still the responsibility of whoever started it. Methods that enlarge a position as it loses, such as martingale and grid systems, are easy to automate and can show a long smooth record before a single run of losses takes the account. An automated strategy needs watching, a limit on what it may lose, and a way to switch it off.