Automated trading terms

AI trading bot

An AI trading bot is a system that uses machine learning to derive or adapt its trading rules, as distinct from a conventional bot executing rules written by a human.

The genuine version applies statistical learning to feature sets and produces a model whose logic is not explicitly written. The marketing version is a conventional rule-based EA with the word attached, and the two are extremely difficult to tell apart from outside.

Machine learning does not solve the fundamental problem and arguably worsens it. A model with thousands of parameters fitted to a few years of price data is the most efficient overfitting machine ever built, and financial time series have a low signal-to-noise ratio compared with the domains where these methods excel.

What adaptivity can legitimately mean is narrower: tracking which of a fixed set of strategies is currently working and weighting toward it. That is what our engine does, using measured win rates with confidence bounds, and we call it a tournament rather than AI because that is what it is. See do AI trading bots actually work.

Algorithmic trading

Also called: Algo trading, Automated trading.

Algorithmic trading is executing trades according to predefined rules encoded in software, removing discretionary judgement from entry, exit and sizing decisions.

The defining property is that the rules are fixed in advance and applied identically every time. This eliminates the largest source of retail underperformance — inconsistent execution of a plan — and replaces it with a different problem, which is that a bad rule is now applied perfectly and tirelessly.

It is not a synonym for high-frequency trading, and it does not require artificial intelligence. Most retail algorithmic trading is a modest set of conditional rules running on a desktop platform, which is exactly what an Expert Advisor is.

The honest framing of the benefit is narrow but real: an algorithm will take the trade at 3am, will not widen a stop because it feels confident, and will stop for the day when the loss limit is reached. Whether that produces profit depends entirely on whether the rules have an edge — see are trading bots profitable.

Backtesting

Backtesting runs a strategy against historical data to estimate how it would have performed, and it is the weakest form of evidence that is still worth having.

Every backtest is a simulation with assumptions baked in — about fills, about spread, about intrabar movement, about which data the strategy was allowed to see. Those assumptions determine the result at least as much as the strategy does, which is why two honest backtests of the same rules can disagree substantially.

The characteristic failures are lookahead bias (using information that was not available at the time), survivorship bias, and overfitting. The first two produce results that are simply wrong; the third produces results that are correct about the past and useless about the future.

A backtest earns trust through what surrounds it: a held-out period, published costs, a stated sample size, and a record of what failed as well as what worked. Ours is published with two separate periods, 27.4% and 13.2% drawdown, and a public list of the 24 strategies that were cut and why.

Copy trading

Also called: Mirror trading.

Copy trading automatically reproduces another account's trades on your own, sized to your balance, so you take the same positions as the account you follow without deciding anything yourself.

A provider trades their own live account. Software watches that account and opens a proportional position on every subscriber account, usually within a second or two. You are not buying a strategy or receiving advice — you are duplicating someone else's executions, whatever they turn out to be.

Sizing is proportional rather than identical. If the provider risks 2% of their equity and your account is a tenth the size, you get a position a tenth as large risking the same 2%. That is the mechanism most people misunderstand: copy trading does not reduce risk, it reproduces it exactly. A provider who blows up takes their subscribers with them, at the same percentage loss.

The genuine trade-off against running a bot is control. A bot runs rules you can read, on your machine, with settings you own. Copy trading runs a person's judgement you cannot inspect, on their schedule, and you can stop it but never steer it. Which is better depends entirely on whether you want to understand what is happening to your money.

Forward testing

Also called: Paper trading, Demo testing.

Forward testing runs a strategy on live incoming data without risking real money, providing evidence that is far stronger than a backtest and slower to obtain.

Because the data has never been seen, forward testing cannot be overfitted by construction. It also captures the things a backtest cannot: real spread variation, real slippage, real broker behaviour and real gaps in the data feed.

Its cost is time. A strategy taking ten trades a month needs the better part of a year to accumulate a sample worth interpreting, and there is no way to accelerate that. Vendors who advertise a strategy weeks after building it are not showing you forward-test evidence, whatever the label says.

Demo accounts are close to live but not identical — fills are often more generous and slippage more forgiving. The gap is usually small on liquid instruments and can be substantial on gold during volatile periods, so a strategy that is marginal on demo should be assumed worse live.

Latency

Latency is the delay between a trading decision and its execution at the broker, comprising network time, platform processing and the broker's own handling.

For retail automated trading the total is typically tens to low hundreds of milliseconds. Whether that matters depends entirely on the strategy: a system entering on a bar close has effectively unlimited time, while one trying to capture a fast reaction does not.

Latency contributes to slippage but is not the same thing. Slippage is the price difference; latency is one of its causes, alongside spread widening and thin liquidity at the moment of execution.

It is heavily marketed by VPS providers and is genuinely the less important of the two benefits they offer. For a strategy scanning every ten seconds and entering on confirmed structure, the difference between 5ms and 50ms is not measurable in the results — uptime is what changes outcomes.

Market regime

A market regime is the prevailing behavioural state of an instrument — trending, ranging or mixed — which determines which category of strategy is likely to work.

The premise of regime detection is that no single strategy works in all conditions, and that the conditions are identifiable in advance often enough to be worth acting on. Trend strategies profit in trends and bleed in ranges; mean-reversion does the reverse.

Detection is imperfect and always lagging, because a regime is only visible once it has been in place for a while. Any regime classifier is therefore describing recent conditions and betting they persist, which is a real assumption rather than a free lunch.

Our engine classifies conditions as trending, choppy or mixed and enables a different strategy set in each. Trending and mixed ship enabled; choppy ships disabled, because in testing we did not have a strategy set that earned money in a range. Disabling a whole regime rather than pretending otherwise is the more useful decision — see range.

Monte Carlo simulation

Monte Carlo simulation reshuffles or resamples a strategy's trade sequence thousands of times to show the range of outcomes its edge could plausibly have produced.

A backtest shows one sequence of trades — the one that happened. Reordering those same trades randomly produces a distribution of equity curves and drawdowns, which reveals how much of the observed result depended on the specific order rather than on the edge.

The output most worth reading is the drawdown distribution. A strategy whose historical maximum drawdown was 20% may show a 5th-percentile outcome of 35% under reshuffling, and that larger figure is the more realistic planning number.

The technique assumes trades are independent, which is imperfect — losing streaks cluster in real markets because regimes persist. That makes the simulated distribution slightly optimistic, so treat it as a floor on the possible pain rather than a ceiling.

A Monte Carlo run is how we worked out how many trades it takes before a track record means anything.

MQL5 Signals

MQL5 Signals is MetaQuotes' built-in copy trading service, where a provider registers a live MetaTrader account and subscribers mirror its trades directly from inside their own terminal.

Because it mirrors an account rather than distributing a file, the strategy behind the trades is invisible and irrelevant to the service. A provider can be trading manually, running an Expert Advisor, or driving the account from an external program in any language — MQL5 only ever sees the executed positions.

The public statistics page is the useful part. Growth, maximum drawdown, trade count, account age and the equity chart are computed by the platform from the broker feed rather than supplied by the provider, which makes them considerably harder to dress up than a self-published screenshot. Subscription revenue is split with the platform.

Registration requires seller status with identity verification, a real account rather than a demo, and leverage at or below 1:500. One trading account maps to exactly one signal, so a provider cannot quietly run several attempts and publish whichever survived — a constraint that does more for subscriber protection than most disclosure rules.

Myfxbook

Myfxbook is a third-party analytics service that connects to a trading account read-only and publishes its performance, and it is the de facto standard for showing a forex track record.

It connects through the investor password or a broker API and then computes the statistics itself from the trade history. Because the numbers are derived rather than declared, a Myfxbook widget is meaningfully stronger evidence than a screenshot of a terminal — which anyone can edit in a browser in about thirty seconds.

Stronger is not the same as sufficient. Myfxbook reports what the connected account did; it does not certify that the account is real money, that it is the only account the vendor ran, or that the strategy shown is the one being sold. All three of those gaps are routinely exploited, and none of them require faking anything.

The checks worth doing take about a minute: confirm the Verified badges for both the track record and the trading privileges, look at account age and total trades rather than the headline percentage, read the gain figure knowing how deposits distort it, and check deposit load to see how much of the equity was actually at risk.

News filter

Also called: News shield.

A news filter suspends automated trading around scheduled high-impact economic releases, when spreads widen, liquidity thins and price moves discontinuously.

The problem it addresses is specific: during a major release the market is not doing anything a technical strategy was designed for. Spreads can multiply, slippage becomes severe, and stops are frequently filled far from their level. Signals generated in those minutes are not signals.

Filters typically read an economic calendar and block trading for a window either side of anything marked high impact. The window matters — too narrow and you trade into the aftershock, too wide and you sit out a large share of the session, since gold reacts to a great deal of US data.

Ours reads the high-impact calendar and pauses new entries around those events. It is worth being clear about what this does not do: positions already open are still exposed, and an unscheduled headline arrives with no warning at all. A news filter reduces a known risk, it does not remove event risk.

What gold actually did around US jobs reports is in news trading on gold: what 11 NFP releases did.

Out-of-sample testing

Out-of-sample testing evaluates a strategy on data deliberately withheld during development, to check whether the apparent edge survives contact with unseen conditions.

The procedure is to split the available history, develop and refine on one part, then test once on the other. Testing on the held-out data more than once destroys its value — each look leaks information, and after a few iterations the out-of-sample period has quietly become part of the training set.

A large gap between in-sample and out-of-sample performance is the signature of overfitting. Some degradation is normal and expected; a strategy that was outstanding in-sample and merely breaks even out-of-sample was fitted, not discovered.

Our shipping configuration was developed on January to July 2026 and then tested on August to December 2025, unseen. It returned +$154 and was positive in three months of five — so the profit generalised, but "green every month" did not. That distinction is recorded in the strategy documentation rather than being quietly dropped.

Overfitting

Also called: Curve fitting.

Overfitting is tuning a strategy so closely to historical data that it captures the noise of that specific period rather than any repeatable market behaviour.

The mechanism is straightforward. Given enough parameters and enough attempts, any dataset can be fitted almost perfectly — and the better the fit to one period, the less likely it is to describe anything general. The resulting backtest looks superb precisely because it is wrong.

It is rarely deliberate. It happens through ordinary iteration: adjust a threshold, the result improves, keep the change. Each individual decision is reasonable and the cumulative effect is a system shaped entirely by one sample of history.

Two defences actually work. Hold data back and test once, and prefer fewer parameters over more. We hit this directly — an earlier version scaled reward-to-risk by market regime, which improved in-sample results and turned out to be fitting noise. Removing the feature and using a single 2.0 ratio everywhere made the system better by making it simpler.

Signal provider

A signal provider publishes their trades for others to follow, either as notifications you act on manually or as an automated feed that executes on subscriber accounts.

The two delivery modes are not equivalent and the distinction matters more than the strategy does. A manual signal — a Telegram message, an email — requires you to be awake, at a terminal, and fast. By the time you read "buy gold at 4012" the price is somewhere else, and the difference between the published entry and your actual fill is the entire margin on most short-term strategies.

Automated delivery removes that gap but replaces it with a dependency: the provider's account, the copier infrastructure and your broker all have to work simultaneously. On MQL5 Signals that plumbing is run by the platform, which is more reliable than a Telegram channel and considerably less flexible.

Assess a provider the way you would assess any system — sample size first, then drawdown, then whether the record is verified. A provider unwilling to show a read-only account is asking you to take performance on trust, which is the one thing that should never be extended in this market.

Social trading

Social trading is copy trading presented as a network — leaderboards, follower counts, public profiles and comments — where discovery happens through popularity rather than through analysis.

Mechanically it is the same thing as copy trading. The difference is the discovery layer: instead of evaluating a track record, you browse a ranked list, and the ranking is doing the persuading.

That layer creates a specific distortion. Leaderboards sort by recent return, and the fastest way to the top of a return-sorted list is enormous risk that has not yet failed. The visible winners are therefore systematically the highest-risk providers in the pool, not the best ones. A provider running 40% drawdowns outranks one running 8% right up until the month they do not.

Read the ranking as a survivorship artefact. The useful columns are the ones the interface makes small — maximum drawdown, account age, deposit load and how many losing months are in the record. A two-month account at the top of a leaderboard has told you nothing except that two months went well.

Strategy tournament

A strategy tournament is a selection mechanism where multiple strategies generate candidate signals simultaneously and only the highest-scoring one is executed.

Rather than running strategies in parallel and taking every signal, a tournament collects all valid signals on each scan and executes exactly one. This bounds risk — several strategies agreeing does not produce several positions — and it forces an explicit ranking rule.

The ranking rule is where the design lives. Ours scores each signal on the Wilson lower bound of that strategy's measured win rate, plus a base priority and a timeframe bonus. Using the lower bound rather than the raw win rate is what prevents a strategy that won its only two trades from outranking one with a long record.

Signals are also deduplicated by strategy type before scoring, which surfaced something useful: several distinct-sounding strategies in our original set were identifying the same setups. That redundancy is invisible when strategies run independently, and it is why three of the four we ship come from different families rather than being variations on one idea.

Trade copier

A trade copier replicates trades from one account to one or more others, either between local terminals or across the internet between different brokers.

Copiers are used for managing several accounts from one signal source, for copy-trading services, and by traders running the same system across multiple prop firm accounts. Position sizes are scaled by balance or by a fixed ratio.

Every copier introduces latency and therefore slippage between master and slave, and the effect is worst on exactly the fast entries where timing matters most. A strategy with a small edge can have that edge consumed entirely by copier slippage across several hops.

Prop firms often restrict or prohibit copying between accounts, particularly copying the same trades across multiple funded accounts, and violations are treated seriously. Check the specific firm's rules rather than assuming — see prop firm rules.

Verified track record

Also called: Verified results.

A verified track record is trading performance published by a third party that reads the account directly, rather than performance reported by the person selling the product.

Verification answers exactly one question: did these trades happen on this account. That is genuinely valuable, because the alternative — a screenshot, a spreadsheet, a video — answers nothing at all. It is also a narrower guarantee than the word suggests, and the gap between the two is where most of this market operates.

What a verified badge does not tell you: whether the account is live or demo, whether the vendor ran fifty accounts and published the survivor, whether the balance is real money or a rounding error, whether the strategy shown is the strategy sold, or whether the settings that produced it are the settings you will be given. Every one of those is compatible with a fully verified record.

Treat verification as the entry requirement rather than the evidence. Unverified results should end the conversation; verified results start it. What follows is the actual assessment — sample size, maximum drawdown, how long the account has run, and whether the losing periods are shown or the record conveniently begins after them.

The questions to put to any bot vendor, including us, are in how to choose a gold trading bot.

VPS

Also called: Virtual private server, Trading VPS, Forex VPS, VPS trading, VPS trade.

A VPS, or virtual private server, is a rented Windows machine in a data centre that stays on all the time. Traders use one to run MetaTrader 4 or 5 and their Expert Advisors around the clock, so trading does not stop when their own computer sleeps, loses its connection or restarts.

Read the full entry: VPS →

Walk-forward analysis

Walk-forward analysis repeatedly fits a strategy on a rolling window of data and tests it on the period immediately after, so every result is out-of-sample.

Rather than one train/test split, it uses many: fit on months one to six, test on month seven, roll forward, repeat. The concatenated test results form an equity curve made entirely of out-of-sample performance, which is a far more honest picture than a single held-out period.

It also answers a question a simple split cannot — whether the strategy needs reoptimising to keep working, and how quickly its parameters go stale. A system requiring frequent refitting is a warning sign regardless of how good the walk-forward result looks.

The main limitation is data hunger. Each window consumes history, so a meaningful walk-forward needs years of it, and low-frequency strategies produce too few trades per window for the individual results to mean much. This is a real constraint for a strategy taking ten trades a month, and pretending otherwise would misrepresent what the method can deliver here. See the walk-forward guide.