Last Tuesday, someone in the Newton community posted a screenshot. A strategy title claimed, "Deep neural network driven, 85% annualized." The image showed a cool-looking neural network architecture diagram, and the description was stuffed with buzzwords like "Transformer" and "attention mechanisms." Underneath, there were seven or eight comments saying things like "Where can I get the address?" and "Big shot, please take me along." I didn’t follow it. I first pulled up that strategy’s on-chain records and studied them for ten minutes.
The result made me laugh out loud.
This so-called "deep learning strategy" has only two trigger conditions for all trading signals within three months: go long when EMA12 crosses above EMA26, and go short when it crosses below. Golden cross and death cross—textbook-level moving-average strategy. It has nothing to do with neural networks. The poster took a first-week assignment from an undergraduate quant course, wrapped it in an AI costume, slapped a 3% management fee on it, and somehow still got a dozen-plus copycats in the market.
After that day, I began systematically digging into the “AI strategies” in the market. Every time I saw an AI tag, I went on-chain to reconcile the records and slowly summarized a set of troubleshooting methods that can work on Newton.
First, look at the signal cadence. When a real AI model makes decisions, the signal distribution is usually uneven—sometimes three or four trades in a day, sometimes none for a week. If a strategy’s signals trigger with extreme regularity—fixed time, fixed frequency every day—then it’s likely a scheduled script. I dug into an “AI high-frequency strategy” and found that it sends a signal exactly at the top of every hour, more reliably than a clock.
Second, look at the data sources. The intro says “combined with multi-factor models,” but the on-chain call records only pull price data—no sentiment indicators, no on-chain capital flow, and no volatility index. Where is the multi-factor part reflected? It’s in the PPT of the introduction. In Newton sandbox execution logs, traces of calls to external data remain—this can’t be faked.
Third, look at behavior during drawdowns. During loss periods, a strategy tends to be much more honest about its performance than it is in profitable periods. I watched one strategy: in a choppy market, it executed an eight-time run of golden cross buys followed by dead cross sells, with almost the same intervals between trades. Mechanical rules don’t adjust based on market conditions—real AI, on the other hand, would show some kind of parameter self-adaptation after a streak of losses.
Fourth, check the publisher’s history of failed drafts. Newton’s on-chain records can’t be deleted. Publishers can delist a strategy, but anything they ran before will persist. I’ve come across people who swapped four strategies in three months: three turned out to lose and were taken down, and they’re now selling the fifth. On-chain records don’t forget—they help you remember.
Finally, run it with small money in real trading. My current fixed approach is to follow it for two weeks with small position sizes. I don’t focus on the return rate; I focus on whether the strategy’s words and actions match. After two weeks, if the execution records line up with the description in the intro, then I consider adding to the position.
After filtering with these methods, about 70% of the “looks great” strategies on the Newton market will reveal their true colors. But the remaining 30% is often worth following. The value of on-chain transparency is exactly this: it doesn’t directly tell you which strategies are good, but it removes the toolkit from people who are just making up numbers. On other platforms you can fool others with PPT, but on Newton, every trade you execute helps keep the accounting. That’s probably the core reason I continue to stay in this market.@NewtonProtocol #Newt $NEWT
