5 min read · 1092 words
We ran the AI side of this comparison for three cohorts and it lost money. Measured win rate 40%, against the 57.4% the exit structure needed. That result sits at the front because it changes how the rest of the page should be read.
TA is not inherently wrong — its patterns reflect real behavioral dynamics of market participants. But it has structural limitations that are especially pronounced in crypto:
Every TA indicator — RSI, MACD, moving averages, Bollinger Bands — is computed from past price data. By definition, a signal that fires when the 50-day MA crosses the 200-day MA is firing after a significant trend has already established itself. In markets that move 10% in a day, lagging indicators are frequently useless for entry timing.
Show ten experienced TA traders the same chart, and you will often get five different pattern identifications and three different trade recommendations. “Ascending triangle” or “rising wedge”? “Cup and handle” or “failed breakout”? The subjectivity is not a feature — it means TA signals are largely unfalsifiable. Any outcome can be explained by TA in hindsight, which makes backtesting TA strategies much harder than it appears.
TA only sees what the market agreed on. It cannot see that a major exchange just had a hack, that a regulatory bill is gaining momentum, that three influential analysts independently came to the same bullish thesis this morning, or that whale wallets have been quietly accumulating for a week. All of that information is invisible to a chart.
TA patterns can be self-fulfilling because enough traders watch them and react to the same signals. But this also means they can be gamed by large players who know exactly where retail stop-losses sit relative to support levels and will temporarily push price through them before reversing.
Honest assessment: Most retail TA traders who have rigorous trade journals will find their TA-based win rate is not significantly above 50% over a large enough sample. The appearance of TA effectiveness is often selection bias — remembering the wins more vividly than the losses.
Three of the four claims usually made for AI crypto signals held up in our build. The fourth is the one that matters.
Breadth held. Where TA reads price, the engine read social, on-chain transfers, news, funding and exchange flows in parallel, clustered them, and scored the cluster. That worked mechanically and still runs.
Coverage held. The scrapers do not sleep, and the moves that happen during the Asian session were seen at the same latency as everything else.
Auditability held. Every signal carried the sources that produced it, which is how the post-mortem was possible at all. Reconstructing why a chart pattern “looked like” a cup and handle six weeks later is not.
Predictive power did not hold. Breadth of input turned out to be a claim about data collection rather than about returns. The conviction score built from all that breadth measured anti-predictive in all three cohorts, meaning the higher-scored entries did no better than the lower-scored ones. Cohort III replayed with fees and slippage set to zero still lost roughly $150. There was no profitable slice at n≥5 by side, source, ticker, entry hour, or conviction bucket.
It settles one thing: multi-source AI synthesis is not automatically better than a chart, because our version of it was measured and was worse than flat. The 40%-vs-57.4% gap is a fact about one engine over three cohorts on one desk, not a proof about the category. It is one more data point than most people arguing this question have.
The trap on both sides is the same and it is arithmetic, not philosophy. Our exits cut winners at roughly +0.7–1% while ATR stops released losers at 3–6.4%. At that asymmetry, breakeven needs 57.4% and no source of entries — chart, model, or tip — was going to reach it. Before comparing entry methods, work out the win rate your own stop and target structure demands. Most of the argument about TA versus AI is being had by people who have never computed that number for their own book.
The engine described here is not a service. The dashboard at /app is a demo on synthetic sample data, and the directional book behind it is closed.