5 min read · 1018 words
Before assessing anyone else’s signal service, here is this desk’s own record. We built a multi-source AI signal engine, paper-traded every signal it produced above threshold, and closed the book.
| Cohort | Result | Why it ended |
|---|---|---|
| I · local | ≈ flat | Indistinguishable from noise at the sample we had |
| II · production | −$33.60 | Headline read +1.2% until the outage inside it was found |
| III · v3.1 | −2.31% NAV | 40% win rate against the 57.4% the exits required |
Three cohorts, none profitable. That is the standard the rest of this page holds other services to, and it is the one thing almost no signal seller will show you.
The economics of a Telegram signal group scam are straightforward and unfortunately highly effective:
The selection bias in signal reporting is the core mechanism. Cherry-picking winners is trivially easy when you send 10 signals per week and only ever reference the ones that worked. Without a complete, auditable trade log, any claimed win rate is meaningless.
More aggressive signal group operators run explicit pump-and-dump schemes: accumulate a low-cap token, then “signal” it to group members who create buying pressure, sell into the rise, and leave subscribers holding bags as the price collapses. This is securities manipulation but extremely common in unregulated crypto markets.
Automated signal bots that connect to exchanges and execute trades are a different category from Telegram signal group tips. They have real problems of their own:
The most fundamental problem with automated bots is that their strategies are almost always backtested on historical data until they look good, then sold as if that performance predicts future results. A strategy that was “optimized” to maximize returns on 2021-2023 data will have parameters specifically chosen to work well on that specific historical period — and will often fail on new data. This is called overfitting, and it is rampant in the retail algorithmic trading bot space.
Most backtest results assume perfect execution at signal price. In reality, market orders on small-cap tokens have slippage, and fast-moving markets often make the advertised entry price unavailable by the time the order executes. A bot with a 5% average win that generates 0.3% average slippage per trade will underperform its backtest significantly.
A bot trained in a bull market will often perform disastrously in a bear market. Momentum strategies that worked brilliantly in 2024 may experience sustained drawdowns in a ranging 2026 market. Without regime detection, automated bots are flying blind.
The one non-negotiable test: Before paying for any signal service or bot, ask for a complete trade log including all losing trades, not screenshots of wins. If they refuse or give you excuses, that is your answer.
Legitimate signal services are distinguished by a small number of non-negotiable characteristics:
Paper-trading every signal above threshold, with no human deciding which ones count, removes the selection bias described above. It does not create an edge. It tells you whether one exists.
The table at the top is what that machinery returned here. Every AI signal over conviction 7 was opened at the entry-zone midpoint and tracked to target, stop, or 72-hour expiry. Nobody could skip a trade after the fact. A replay of cohort III with fees and slippage set to zero still lost about $150, so the losses were not costs — the entries had no edge. There was no profitable slice at n≥5 anywhere in it: not by side, not by source, not by ticker, not by hour of entry, and not by conviction bucket. The conviction score, the headline number the whole product was built around, measured anti-predictive in all three cohorts.
Cohort II is the more instructive failure. It reported +1.2% for weeks. The trader daemon had been down for roughly two weeks, two positions turned into accidental 131-hour holds, and both closed on one restart tick into a 21% SOL rally. Remove the outage and the cohort was negative. A track record that is automatic is still only as good as the infrastructure underneath it, and it took forensics rather than accounting to find that.
So when you apply the trade-log test to a signal seller, apply the second one too: ask what they found when a result went against them, and what they did about it. This desk closed the book. The write-ups are on the front page, cohort by cohort, with the numbers that closed each one.