How frustration with our own bad calls turned into a system that sticks to the rules when people don't. The road to the first beta — and what comes next. Without glossing over the detours.
Why this bot exists at all.
Discipline under pressure isn't reliable.
A plan that worked, two red days, and the stop gets moved. Not because the market changed, but because losing hurts. The idea didn't come from spotting a gap in the market — it came from realising that discipline isn't reliable under pressure.
Rigid rules only push the problem further ahead.
The first scripts were simple: fixed thresholds, fixed conditions. In the phase they were born from, they did well. When the market turned, they kept running — just wrongly. Rigid rules don't solve the problem, they postpone it.
“The idea didn't come from spotting a gap in the market — it came from realising that discipline isn't reliable under pressure.”
People and data, before any model ran.
Three perspectives that correct each other.
A trader who knows where strategies break in practice. An ML specialist who knows when a model is fooling itself. A developer who makes both usable. No investor money, no deadline — but the freedom to throw approaches away.
On bad data every model convinces, but only on paper.
Before the first model came months of cleanup: gaps in histories, shifted timestamps, prices from periods without real liquidity. A model that convinces on bad data only convinces on paper.
What was tested, discarded and kept.
A backtest proves nothing. The time after it does.
Several model families competed — from simple decision trees to networks built for time series. The most interesting finding wasn't which one won, but how many shone in backtests and fell apart outside their training window. Textbook overfitting, visible only in walk-forward validation.
No majority, no signal.
The obvious move would have been to pick the best model. Except none was best across all market phases — each had regimes where it beat the others, and regimes where it got it wrong. What remained are three with different views: trend, momentum, volatility. They vote on every setup, their votes carry different weights, and a signal only forms above a threshold. Whichever keeps getting it wrong loses weight.
Confirmation is not the same as repetition.
The first instinct is to collect indicators. The second is to remove them again. Many measure the same thing with different formulas, faking confirmation where there's only repetition. What remained is a small set that genuinely sees different things.
Volume shows where a move has substance.
A price line tells you where price was. It doesn't tell you where volume actually sat or which zones were contested. Market profile added that structure: the point of control as the most-traded level, the value area as the range holding the bulk of the volume. With it, a sense of which moves have substance and which are just noise.
“Most of the approaches we tested never made it into the product.”
The layer with a veto.
The model may be wrong. The risk limit may not.
A signal is allowed to be wrong; a risk limit isn't. Take-profit, stop-loss and staged entries via DCA are built in, plus a drawdown limit per engine and a command that closes everything at once. The layer holds a veto: it can block any signal, while the model cannot override it.
Where we stand today.
Gold first, because that's where you see most clearly what the engine does.
The engine runs live on XAU/USD — limited seats, trades published as they happen, access only after approval. Gold first, because its sessions are clearly structured and behaviour is easier to judge than in the round-the-clock crypto markets.
What comes next, without dates.
A second pair of eyes on every rule.
Another trader takes over fine-tuning the engine: sharpening signals, re-testing strategies against new market phases and double-checking the risk rules. A second pair of eyes on what one person has carried so far.
More markets, the same engine, the same account.
Major currency pairs, Nasdaq and DAX join the existing engine — through the same VT account, with nothing new to open. Traded far more conservatively than gold: fewer positions, longer holding periods.
Crypto never sleeps. Neither does the risk layer.
Round-the-clock crypto makes different demands: no session close, no trading break, thinner order books at weekends. Risk and execution have to prove themselves across several market phases first. Runs on its own exchange, not through VT.
Live when it convinces. Not on a calendar.
Altcoin breadth comes last because it's the hardest: many pairs, erratic liquidity, correlations that flip without warning. It goes live once it has convinced across several market phases — not on a calendar date.
This roadmap describes how the system was built, not how it performs. It deliberately contains no return promises. Current figures are in the track record; every live trade can be followed in the Telegram channels.
Four principles that came out of the failed attempts — and haven't changed since.
Every approach has to hold up outside its training window — on phases it has never seen. A model that only works in hindsight gets dropped, however good the curve looks.
Not more indicators, but ones that measure different things. Plus the question of where volume actually sat — that's what separates moves with substance from plain noise.
Stops, position sizes and drawdown limits aren't options you can switch off. They sit above the model and apply even when a signal sounds convincing.
Entry, exit, result — all of it runs publicly in the Telegram channels. If you trust an automated system with capital, you should be able to see what it does. On bad days too.
No investor money, no outside deadline. Instead, perspectives that correct each other — soon one more.
Brings the practice: which setups hold up in a real market, where they break, and what level of risk is bearable over months.
Owns model selection and validation — and the uncomfortable job of discarding approaches that merely confirm themselves.
Builds everything users actually see: console, reports, Telegram integration and the paths in between.
Takes on signal fine-tuning, re-tests against new market phases and cross-checks the risk rules — a second pair of eyes on what one person has carried.
We deliberately operate without names or faces. Anyone publicly associated with trading systems becomes a target — for social engineering as much as for copycats. The work should stand on its own: every trade is publicly verifiable, even when we aren't.
Watch the bots trade first. The live channels are open, free and without obligation.
Bot news, weekly reports and updates all run in the main channel. No newsletter, no emails.
Open main channelThe information presented on this page is provided for general information purposes only and does not constitute investment advice. Trading leveraged products, crypto assets and FX pairs is highly speculative and carries substantial risk — up to and including the total loss of the capital employed. Past performance is not a reliable indicator of future results. ATX BOT does not hold client funds; all positions are held via API in the user's own broker account. ATX BOT is trading software.