How to Check a Trading Algorithm Before Buying: Full Investor Checklist

Why you can not choose a trading algorithm only for return

Profitability without context says almost nothing about the quality of the system. The algorithm can show +100% in a few months, but do so through a risk that does not suit the investor: a huge position size, no stop loss, averaging against movement or a series of trades with a critical open drawdown.

When checking a trading algorithm, not only profit is important, but also the path to this profit. You need to look at the maximum drawdown of the trading algorithm, the number of trades, the length of the test period, the stability over the years, the average trade and how the system behaves in adverse market conditions.

One successful series of deals does not prove the sustainability of the strategy. Short statistics may not include a crisis, a side market, a sharp increase in volatility, or a period of low activity. Therefore, the question of “how to choose a trading algorithm” should always begin with an assessment of risk, not with a search for the highest returns.

Particularly dangerous proposals, where the yield is shown separately from the size of the account, the volume of the position and the drawdown on funds. For example, a algorithm could earn many percent on a small account just because it used an excessive shoulder. It is more important for the investor to understand how much capital was at risk at the worst moment and whether it was psychologically and financially possible to sustain such a period.

How to check the results of the backtest of a trading algorithm

Trading algorithm backtesting is a test of strategy on historical data. It shows how an algorithm might have traded in the past under given rules of entry, exit, and money management. But a beautiful capital chart is not in itself proof of reliability.

When analyzing the report, you need to check not only the final profit. The duration of testing, the number of transactions, the maximum drawdown, the average profit or loss per transaction, the profit factor, the profit and drawdown ratio, the stability of results by year and the presence of periods of stagnation are important.

• testing period and number of transactions;
• profit and maximum drawdown;
• average transaction after expenses;
• profit factor and profit/drawdown ratio;
• stability of results by year;
• long periods of stagnation;
• uniformity of the capital curve;
• Strategy in times of crisis.

It is desirable that the testing period include trend markets, side markets, periods of high volatility, crisis areas and periods of low activity. If the algorithm was only tested on a short, handy piece, the algorithm’s check remains incomplete.

Separately, it is worth looking at the distribution of profits. If almost all of the profits are made in one month or several large trades, and the rest of the story looks neutral or unprofitable, this result cannot be considered sustainable without additional analysis. A good report should show not only the total figure, but also the nature of the trade: how often the algorithm opens trades, how long the position lasts, how deep the series of losses is, and how quickly the system recovers.

It is also important to compare the size of the average transaction with typical expenses. If the average trade is too small, the algorithm may be sensitive to spread changes, server latency, or execution quality. In the test, such a strategy looks neat, but in real trading, even a slight deterioration in conditions can take away the advantage.

Quality of historical data

The result of the backtest directly depends on the quality of quotes. Minute data, tick data, and artificially simulated movements inside a bar can produce different results, especially if the strategy works on low timeframes, uses stop orders close to price, or depends on the exact moment of entry.

It is important to understand how the movement inside the candle is modeled. If a algorithm enters and exits inside the same bar, rough simulations can create trades that would not be possible in reality or would be executed at a different price. Skips in quotes, history errors and incorrect instrument specification also distort the result.

Percentage of model quality alone does not guarantee the reliability of the test. You need to look at the data source, data type, tool specification, spread size, commission, time zone and whether the test can be repeated on an independent set of quotes.

For CFDs and foreign exchange instruments, the problem of single broker data is particularly important. Quotes, spreads and trading session times may vary, so an algorithm who looks perfect on one broker is not required to repeat the result on another. For futures, there is another task: to glue contracts with different expiration dates correctly so that the history does not contain artificial leaps.

If the seller does not disclose the source of the data or responds only with the generic phrase “99% model quality”, it is worth asking for more details. It is useful for the buyer to know on which symbol the test was conducted, what trading conditions were used, whether there was a fixed or variable spread, and whether the specification of the instrument matches the one on which the actual trading is planned.

Whether commissions, spread and slippage are taken into account

Without trading costs, the result can be greatly overstated. The trading algorithm should be tested taking into account the brokerage commission, typical spread, spread expansion, slippage, position transfer fees and, if we are talking about futures, exchange and clearing fees.

Commissions should be considered both at the entrance and at the exit of the position. Be especially careful with algorithms that show frequent trading with a small average trade. In such a system, even a small cost error can completely change the outcome.

The spread expansion cannot be ignored either. In calm hours, the spread can be small, but before news, at the opening of the market or during periods of low liquidity, it can increase sharply. If the algorithm is actively trading at such moments, a test with a constant narrow spread can look much better than reality.

Slippage is especially important for algorithms that use market orders, stop orders, and quick entries after a breakdown. The buyer should understand whether slippage was included in the test and whether the seller has statistics of actual performance. Without this, the verification of the trading algorithm remains incomplete.

Checking the algorithm on data beyond optimization

Data beyond optimization is a piece of history that wasn’t used to match the algorithm’s parameters. In simple terms, this is the “unfamiliar” part of the past for the algorithm, which checks whether the strategy was too precisely adjusted for one convenient period.

If good results are only in the optimization area, this is an alarming signal. This situation may indicate over-optimization: parameters are ideally suited to the past fragment, but lose stability when the market changes. On data unknown to the algorithm, the performance usually deteriorates, and this is normal. It is important that the deterioration does not destroy the trading logic itself.

The non-optimization test helps to assess whether the positive structure of transactions is maintained, whether the drawdown remains within reasonable limits, and whether the capital curve turns into a random set of successful and unsuccessful periods.

It is normal if the results on data outside optimization are more modest than on the site of selection of parameters. This is natural: the algorithm no longer “saw” this data during the setup. An alarming sign is that when profits completely disappear, the drawdown increases sharply, the number of loss-making series increases, and the logic ceases to look stable.

The buyer does not need to demand the same numbers in all parts of history. It is more important that the strategy maintains a clear pattern of behavior. If the system was trendy, it may work worse in flat, but it should not destroy the score. If the algorithm is designed to return to the mean, he must have a clear mechanism to limit the risk in strong directional motion.

Forward Test of Trade Algorithm

Backtest looks to the past. The algorithm’s forward test shows how the strategy behaves after the development is completed, when new data is no longer used to match parameters. Real trading adds another layer of verification: actual execution, spread, slippage, and psychologically important account transparency.

The forward test should be carried out without changing the parameters retroactively, for a sufficiently long period, with real trading costs and preferably on an independent account. Demographics are useful for initial logic testing, but they don’t always reflect real execution, especially for scalpers, news algorithms, and small-medium-deal strategies.

It is important to distinguish an honest forward test from a constant manual fit. If, after each series of losses, the seller changes the parameters and then shows the updated chart as a “forward”, such statistics become meaningless. A forward test is useful precisely because it shows the behavior of a predefined system under new conditions.

The demo test can be used as a filter: check whether trades are opened according to the declared logic, whether there are technical errors, whether the frequency of trading coincides with the description. But before a serious purchase or use on a large account, it is advisable to see at least some of the statistics on the real execution, because this is where the costs, delays and restrictions of the broker manifest themselves.

Verification of real account statistics

The real statistics of the trading algorithm is more important than the advertising screenshot of the terminal. The screenshot is easy to take out of context or edit. Account monitoring should allow you to check the date of the start of trading, current profit, maximum drawdown, history of transactions, deposits and withdrawals, account type and whether real or demonstration mode is used.

The seller should request the duration of trading, the number of transactions, the size of leverage, open and closed drawdown, data on the broker, account type and compliance of transactions with the declared strategy. It is important to understand whether the increase in risk is hidden: for example, the growth of the lot after losses or the change in parameters after an unsuccessful series.

Real account monitoring should also be read carefully. Replenishing an account can visually smooth the drawdown, withdrawing funds can change percentages, and a too short history does not show behavior in different market regimes. It is useful to look not only at the total profit, but also at the sequence of transactions, periods of inactivity, maximum open load and position size relative to capital.

If the seller shows a real account but hides the history of transactions, the broker, the type of account or the drawdown of funds, confidence in the statistics decreases. Transparency does not mean that the system is risk-free. It means that the buyer can independently assess the risk and decide whether it is suitable for his capital.

Maximum drawdown and real risk

The maximum drawdown shows how much capital has declined from the previous high. This is one of the key parameters when buying a trading algorithm, but it cannot be taken as a guaranteed limit for future losses.

You need to distinguish between a drawdown on the balance and a drawdown on the account. The closed drawdown reflects losses already recorded. An open drawdown shows current floating losses on unclosed positions and can be more dangerous if the algorithm holds losing trades for a long time or averages against movement.

If the maximum drawdown in the test was 15%, the buyer should not assume that in real trading it will never exceed 15%. Market conditions change, performance may differ, and increasing the size of a position directly increases money risk.

You can read more about the systemic approach to risk in the article about Risk management in algorithmic trading.

The practical approach is to determine in advance the level of loss after which the trade should be reduced or stopped for analysis. Such a limit does not make a strategy profitable, but it helps not to make decisions in the moment of emotional pressure. If the seller cannot explain what happens after a series of losses, the risk of capital management remains unclear.

How to recognize martingale and grid trading

Martingale is an approach in which the next position increases after a loss. Averaging – adding new trades against price movements to improve the average entry price. A trading grid is a series of orders at different price levels. These methods are not necessarily fraud, but their risk must be fully disclosed to the buyer.

Such systems can show a smooth profit for a long time, because small market pullbacks close a series of trades in the plus. But one strong directional series can lead to a critical drawdown, especially if there is no fixed stop loss and the position size increases.

• Increased position after loss;
• several unidirectional transactions with increasing volume;
• the absence of a fixed stop loss;
• a very high percentage of profitable transactions;
• Small average profits and rare huge losses
• long-term retention of unprofitable positions;
• The promise of an almost complete absence of loss-making deals.

Dependence of the trading algorithm on the broker

The dependence of the algorithm on the broker arises due to differences in quotations, spread, commission, execution speed, slippage, minimum position size, volume step change, trading session time, instrument specification, server location and time zone.

Scalpers, news algorithms, strategies with a small average trade, algorithms on low timeframes and systems that put pending orders close to the current price are particularly sensitive to the broker. If the seller requires the use of only one unknown broker and does not explain the reason, this is an occasion to ask additional questions.

When choosing between markets, it is also helpful to understand the differences in infrastructure. For example, an article Futures and CFDs for Algorithmic Trading It explains why the same algorithm can behave differently in different environments.

Dependence on the broker is not always a problem in itself. Some strategies do require certain trading conditions: a low spread, stable execution, a specific account type, or a close server location. The problem begins when these requirements are hidden and the buyer is promised the same result in any infrastructure.

Signs of a Bad Trading Algorithm

What questions to ask the seller before buying

1. At what time did the backtest take place?
2. What historical data was used?
3. Are commission, spread and slippage accounted for?
4. Are there results beyond optimization?
5. Was there a forward test?
6. Is there a real account monitoring?
7. What's the maximum drawdown?
8. Is martingale, grid or averaging used?
9. Is there a fixed stop loss?
10. How is the position size calculated?
11. What brokers and account types are the algorithms designed for?
12. How often are the parameters updated?
13. What market conditions are unfavorable?
14. Is it possible to test a algorithm on a demo account?
15. What technical support is provided after the purchase?

Checklist of the buyer of the trading algorithm

Rule: Do not buy a algorithm until at least the backtest, data outside optimization, forward test, drawdown, method of money management and real account statistics are checked.

How Algo Trade Systems Checks Trading Systems

Algo Trade Systems sees trading algorithms as system strategies, not as a set of advertising promises. In the development process, a long history, accounting for commissions, spreads and other costs, verification beyond optimization, forward testing, analysis of drawdown and stability of parameters are important.

If a product is positioned as a trading algorithm without a martingale, this should be reflected in the logic of money management. It is also important to describe the recommended terms of use: platform, account type, market, acceptable risk, and strategy constraints.

It is useful for the buyer to compare a single algorithm not only with other products, but also with other products. portfolioRisk can be spread across multiple algorithmic strategies.

FAQ: Checking a trading algorithm before buying