Algorithmic Trading: a complete guide to the method
Algorithmic trading executes predefined rules mechanically, removing emotion and discretion from decisions.
| Parameter | Value |
|---|---|
| Execution uptime | 99.9%+ |
| Emergency shutdown | Kill-switch |
| Of every trade | Logging |
| Timeframe | Any |
How it actually works
The essence of the algorithmic approach lies not in speed or complexity but in removing the human from the moment of decision. Rules are formulated in advance, programmed unambiguously and executed without exceptions — and it is this discipline, rather than the brilliance of the rules, that usually provides the edge over manual trading.
The unambiguity of the rules has a downside too: the algorithm does exactly what is written, including the author’s mistakes. A strategy fitted to history will confidently trade on useless patterns, and an unnoticed leak of future data into the backtest will turn a losing idea into a spuriously profitable one. That is why, in algorithmic trading, validating rules matters more than inventing them.
A separate danger is over-optimisation. The more parameters there are and the more carefully they are tuned to the past, the prettier the backtest and the worse the behaviour on new data. Working algo systems are usually simple, have few parameters and are tested on data that did not take part in development.
Why this methodology cannot be honestly tested on our data
Algorithmic trading is not a separate strategy but a way of executing any rules, so “testing the methodology” with a single backtest is impossible: the result would depend entirely on which specific rule is plugged in, not on the approach itself. Showing one run (say, a Donchian breakout rule) and passing it off as an assessment of algo trading as a whole would be a substitution of concepts — one could just as easily take a losing rule and “disprove” algorithmic trading. What is always tested is a specific algorithm with fixed rules, not a category.
We deliberately show no backtest here: presenting attractive figures computed on unsuitable data would mislead the reader. An empty space is more honest than an invented result.
Pros and cons
- Fully removes emotion and discretion from execution.
- Any rule can be formalised and tested on history by one method.
- Scales to many instruments simultaneously.
- Executes the author’s mistakes too, without exceptions.
- Vulnerable to over-optimisation: a pretty backtest often fails to work live.
- Requires programming skills and execution infrastructure.
Nuances and pitfalls
The chief enemy of algo trading is overfitting to history. A strategy with a dozen tuned parameters will almost always show a wonderful result on the past and fail on the future, because the parameters describe the randomness of the sample, not regularities of the market. The second frequent failure is look-ahead in the backtest: using data in the signal that, in reality, would only have become known later. Such a leak makes the backtest unattainably profitable, and it is uncovered only on a live account.
Who this methodology suits
For traders with programming skills and a bent for systematic thinking, ready to spend the bulk of the effort not on generating ideas but on honestly validating them. Requires infrastructure and an understanding of statistics.
Frequently asked questions
Do I need to be able to program for algo trading?
Yes, at least at a basic level. Visual strategy builders exist, but they limit both the logic and, more importantly, the ability to test the system honestly. Without code it is hard to run a correct backtest that accounts for costs and avoids look-ahead.
Why does a profitable backtest often fail to work live?
Most often because of over-optimisation: the parameters are tuned to past data and describe its randomness rather than stable regularities. Less often it is because of look-ahead or underestimating costs. Testing on data not used in development, together with realistic costs, filters out most of these illusions.
Is a complex strategy better than a simple one?
Usually the opposite. Every extra parameter increases the risk of fitting to history. Robust algo systems tend to be simple, have few degrees of freedom, and therefore describe noise instead of signal less often.