Before You Pay for a Bot, Ask to See the Losing Week

The request that does the most work on this whole list is one sentence long: show me an exported trade history from a losing week. Before we get to why, the disclosure that has to come first: this blog is published by PraxAI, we sell trading automation software for prop firm accounts, and a list of trading bot red flags written by a company that sells bots is self interested by definition. Read it that way. Every check below works on us too.

The checks are cheap and the mistake is expensive. A trader who just burned an evaluation fee shops in a hurry, and a picture of an account is the weakest evidence in this business. Nothing in it can be verified, and a seller whose only proof is pictures knows that.

There are no company names here, because the same product can be sold honestly by one desk and dishonestly by another. What you can evaluate is the shape of the claim and the answer to a narrow question. For the wider framing, the [guide to choosing an AI trading bot for prop firms](/blog/best-ai-trading-bot-prop-firms-2026) is the main read on the topic. This post is the screening pass before it.

The first flag is visual. A real curve has flat weeks, ugly steps down, and stretches where nothing happens because a filter kept the robot out. A line that rises almost straight for months means the sample is short, the losses were never closed, or this is a fitted backtest with an account label.

Learn to spot the smooth line that rises while a losing position sits open underneath it: closed profit looks perfect, open risk is invisible. Ask for the period plotted as balance and equity together, because the gap between those lines is where hidden drawdown lives, and the limit that ends a prop account is usually measured on equity. If that is new, read [how prop firm account metrics are calculated](/blog/prop-firm-account-metrics-explained) first.

The second flag travels with it: a win rate with no average loser beside it. A system can win nine out of ten trades and lose money, and very high win rates are often bought by holding losers longer than winners.

Demo results and backtests are not worthless, they are a different category of evidence, and the flag is a seller who blurs the two. Demo fills are optimistic by construction: spreads behave, requotes do not happen, and the order that would have missed by half a point gets filled anyway. A backtest is often the strategy graded on the same data used to shape it.

What makes this checkable is methodology. A serious backtest arrives with the date range, the symbol, the spread and commission model, the modelling quality, and a statement about whether the parameters were optimized on that same period. A seller who has that sends it within a day. A seller who does not sends another image, which is why [backtest versus live EA results](/blog/backtest-vs-live-ea-results) diverge so often.

A related flag is a live record with no start date. Live since launch is not a date. Ask when the current version started trading, because a robot reconfigured last month has a track record that begins last month.

The most expensive pattern hides behind vocabulary. Recovery mode, smart averaging, position reinforcement, layered entry, cost basis optimization: the label rotates, the behavior does not. You do not need source code to identify it, because it shows up in the trade history of any losing week.

Two signatures give it away. First, position size that increases after a loss instead of staying flat. Second, several entries stacked in the same direction while price runs against the first, with no stop closing any of them. Either one means risk that compounds, the profile that produces a beautiful curve for months and then one session that removes the account.

It matters more after funding than during an evaluation, which is why the behavior gets its own read: [what martingale and grid EAs do to prop firm accounts](/blog/martingale-grid-ea-prop-firms) covers the mechanics and the rules that catch them. It is also one of the ordinary reasons [funded traders lose the account](/blog/why-funded-traders-lose-the-account) after a strong first month.

Nobody is owed a vendor's source code, and asking for it is not the test. The test is whether the seller can describe the robot at the level of one order: what makes it enter, where the stop goes, how the stop moves, what closes the trade, and what keeps it flat. That contains nothing worth stealing, and a seller who cannot produce it either does not know the logic or is hiding it.

Watch for the reply that swaps technology in for mechanics. Neural network, machine learning and adaptive AI can all be true and can all be filler. The honest version says which part is a trained model and which part is a fixed rule in code, which is the subject of [whether any of these bots are actually AI](/blog/is-any-trading-bot-actually-ai).

The rule side has its own flag: no set files per firm, no rulebook mentioned anywhere. Prop accounts differ in daily loss limit, in maximum drawdown and whether it trails, in minimum trading days, in news and weekend handling, and all of it varies by firm and account type and changes over time, so confirm the current version on your firm's own site. One setting for every account means the robot was never mapped to the constraint it has to survive. A per firm configuration is the visible sign that the mapping happened at all, and [whether bots are allowed at prop firms](/blog/are-trading-bots-allowed-prop-firms) is the check that comes before any of it.

Some flags are not technical. Any promise of income, of passing, or of a monthly percentage is disqualifying on its own, because no software vendor controls the market, your firm's rulebook, your broker's execution, or your hand on the panel. Advertising rules exist around that language, and a seller using it anyway has decided they do not apply.

Scarcity that resets is next. A counter that says nine licenses left and still says nine next week, a countdown that restarts on refresh, a price that has been ending tonight for a month. None of it is fatal alone, but it says the desk is optimized for pressure, and those desks tend to run support the same way.

Then the quiet one: no identifiable support channel before you pay. If the only contact is a web form, if there is no named place where customers talk, if a pre sales question sits unanswered for a working day, that is what support will feel like when your account is near its limit on a Friday.

This is the most efficient question on the list, and it is not about the money. Ask three things at once: what is the refund policy, what disqualifies me from it, and how long does a refund take. Then read the reply for specifics rather than tone.

An honest answer has a window, a set of conditions, and a mechanism, including the uncomfortable conditions. A weak one is warm and vague, redirects to results, or slides into reassurance that you will not need it. A policy that names no condition has probably never been written down.

Pair it with one more: what does support do if the robot fails the evaluation. A real answer is modest and concrete about logs, diagnosis and configuration. The absence of one, replaced by the claim that failure will not happen, is the promised outcome flag wearing different clothes.

Every check above works on PraxAI, so here are our answers. The gold configuration we validate takes one position at a time, no martingale and no grid, verifiable in a trade history rather than on our word. PraxAI GUARD is not AI: it is a set of limits you define, enforced in code, closing a position when your limit is reached and blocking an entry that would breach a ceiling. PraxAI SIZER is a sizing panel for orders you place by hand and never opens a trade itself. The license is 497 dollars once, unlimited accounts, a 7 day money back guarantee, rule updates within 48 hours.

The limits, plainly: MetaTrader 5 is the mature platform, the cTrader cBot is new and in validation with the first customers, and NinjaTrader 8 covers futures, where whether automation is permitted depends on each firm's own policy, so confirm that in writing. We promise nothing about passing, because the outcome is not ours to control, and only 1 to 3 percent of funded traders keep the account long term.

If our answers come out worse than someone else's, buy from someone else. That is the point of a screening list rather than a ranking. Run it on three vendors, keep the replies side by side, and let the pattern decide. When the screening is done and you are choosing between the survivors, the [six questions that cover fit and commercial terms](/blog/how-to-choose-prop-firm-trading-bot) pick up where this list stops. The loudest promise in this category, passing an evaluation in minutes, gets its own teardown in [what HFT expert advisors actually do to a prop account](/blog/hft-ea-prop-firms-reality).