Algorithmic and AI-Assisted Trading: Systematic Rules and How to Test Them

A systematic trader is one whose decisions follow written rules. Given the same market, the rules produce the same action every time, whoever is at the screen. That can be done by hand, with a checklist, or by software. The earlier article on algo trading versus automated trading covers the difference between deciding and executing, so this one concentrates on the rules, a classic example, and how to judge claims about AI.

What a systematic rule set contains

A complete system answers five questions in advance. If any one is left to mood, it is partly discretionary.

1. Universe: which instruments it trades and on which timeframe.
2. Entry: the exact condition that opens a trade.
3. Exit: the stop, the target or the trailing rule, and any time limit.
4. Size: how many units to trade, usually from a fixed percentage of the account risked.
5. Limits: the maximum open risk and the loss at which trading pauses.

A rule is only useful if two people reading it would act the same way. The phrase buy a strong uptrend is not a rule. The phrase buy when the daily close is above the 50-day average and the 14-day RSI crosses back above 40 is one. Writing rules this way also exposes how many of your usual decisions were really judgement calls.

A classic example: Elder's Triple Screen

Alexander Elder developed the Triple Screen in the mid 1980s, and his book Trading for a Living describes it in full. It is a good example of a rule set because each screen has a single job and an indicator that answers it, and it is a clear case of the top-down method in multi-timeframe analysis.

Elder named the screens by images from the sea. The first is the tide, the long-term direction. The second is the wave, the pullback against the tide. The third is the ripple, the small move used to time the entry. Each chart is a clear step shorter than the one above it, for example a weekly chart paired with a daily one, and a daily chart with a 4-hour or 1-hour one.

The key rule is that the first screen has a veto over the others. When the weekly trend is up, the system looks only for buys, and uses daily dips as the chance to enter. When the weekly trend is down, it looks only for sells on rallies. A daily oscillator that says sell is ignored in a weekly uptrend. The MACD and multi-indicator articles explain the tools.

A worked example with invented prices

Suppose a weekly chart of GBP/USD shows the MACD histogram rising, so the tide is up and only buys are allowed. On the daily chart the Stochastic has fallen below 20, which is the usual oversold convention. That is the wave. The prior day's high is 1.2650, and the low of the pullback is 1.2540.

If the order is not triggered, the buy-stop is moved down to just above each new day's high, so the entry only happens when price turns up. If the weekly histogram turns down before the order fills, the order is cancelled. With 112 pips of risk, position size follows from the formula in how much to risk per trade. The system is a framework. It does not promise that its trades win, and the choice of indicators and settings is itself a decision to test.

Where AI fits in

AI-assisted trading usually means machine-learning models used to find patterns in price or other data, or language models used to read news or help write code. Some of that is useful. A model can classify text, flag unusual activity or help you write and check a script faster. None of it changes the basic problem, which is that financial data is noisy, the relationships in it change, and an edge once found tends to fade as others find it.

Regulators have started acting on exaggerated AI claims. On 18 March 2024, the US Securities and Exchange Commission announced settled charges against two investment advisers, Delphia and Global Predictions, for making false and misleading statements about their use of artificial intelligence. They paid civil penalties of $225,000 and $175,000. The lesson for a trader is that a product calling itself AI-powered has not shown that it works.

How to judge a claim about an AI trading tool

Overfitting

A model with many adjustable settings can be tuned until it fits past data almost perfectly, including the random noise. It then looks brilliant on history and fails on new data. The more variations tested, the more likely the best one is simply the luckiest. Bailey, Borwein, Lopez de Prado and Zhu published a method for estimating the probability that a backtested strategy is overfit, because the usual hold-out test is unreliable when many variations have been tried. A simple illustration: if you test 1,000 strategies that are pure coin flips, a handful will look impressive over any single period.

Backtest versus live results

A backtest is a simulation, and it tends to be kinder than reality. It can ignore the spread, slippage and commissions, assume fills at prices that were not available, or use information that was not known at the time, which is called look-ahead bias. A fair question to ask is how the strategy did on trades made with real money, over a period after it was finished.

Survivorship bias

If a test uses only the stocks that exist today, it leaves out the companies that failed or were delisted, which makes strategies look better than they were. The same applies to showing only the best of many models or traders. A track record chosen after the fact from many attempts is not evidence of skill.

A short checklist

• Is the logic explained well enough that you could say why it should work?
• Were the results tested on data the model never saw, and over different market conditions?
• Do the results include realistic spreads, commissions, slippage and financing?
• Are there live results, ideally verified by a third party rather than screenshots?
• How many versions were tried before this one was chosen?
• Does the seller promise returns, call it risk-free or hide how it works? Each is a reason to walk away.

Testing your own rules

The same tests apply to a system you build yourself. Keep the rules few, because each added parameter is another chance to fit noise. Split your data into a period for building and a period you do not touch until the end. Use costs that are slightly worse than you expect. Then trade it at very small size for a meaningful number of trades before trusting it, and compare the live results with the backtest. A large gap means something was missing from the test.

Automation also brings its own risks. A bug, a lost connection or a changed market can repeat a mistake many times before you notice, so a systematic trader sets a daily loss limit and watches the system. Trading with leverage can lose more than the amount deposited, and no rule set, however well tested, changes that.

Sources: Alexander Elder, Trading for a Living, and his Triple Screen system; US Securities and Exchange Commission press release 2024-36 on its first AI-washing enforcement actions; Bailey, Borwein, Lopez de Prado and Zhu, The Probability of Backtest Overfitting, Journal of Computational Finance. Prices in the example are invented.