Why I let AI reverse-engineer two strategies
I have always said that for most people, DCA is the most grounded way to build a long-term position in an asset they believe in. You do not need to trade the range, read leverage signals, or follow every slogan on social media. But the pace of AI development has changed what people ask me. Many of you keep sending examples of automated trading strategies and asking whether AI can turn them into something that fits your risk profile.
I now spend a huge amount of compute every month, running AI agents across models including DeepSeek and Grok. I use them to research, write code, check trades, and test ideas. When I came across two strategies with public trade histories, I gave the records to my AI agents and asked them to reconstruct the rules, verify the trades, and rerun the same starting capital on the same data.
The comparison is simple. For each period, I start with 10,000 USDT in three accounts: one that buys and holds, one that follows strategy 1, and one that follows strategy R. The results I show are historical reconstructions, not promises. Trading always has two outcomes: profit or loss. I am not giving investment advice here. If we cannot accept that, we should not touch any of this.
Strategy 1: trend-following with tranche entries
The first strategy is long-only and works on spot markets. It begins by assessing the main trend. When the trend is strong, the main position rides the move. When the trend weakens, it switches to a smaller position and buys in tranches as price falls. The goal is to bring the average entry price down.
The reconstruction includes a martingale-style multiplier of 1.5 at each level. That means later buy orders get larger, which can lower the average cost during a recovery, but also increases risk. The AI suggested spacing those additions according to volatility rather than using a fixed distance between levels. Once the tranche portion reaches its preset target, it takes profit.
In the parameters I reconstructed, the main exposure could be as high as 90% of the account at the initial entry, with the first martingale tranche around 1% and each subsequent tranche multiplied by 1.5. Those numbers may be approximate, but the structure is clear: small initial additions become progressively larger.
Strategy R: four indicators vote and the position can flip
The second strategy is more active. It can go long or short, so it tries to make money in both up and down moves. Four indicators vote: MACD for trend direction, KD or stochastic for where price sits inside the recent range, RSI for momentum strength, and DMI for directional strength. The order is executed at the next candle's open.
When the signal flips, the strategy first closes the existing position and only then opens the opposite one. Shorting is a risky trade in its own right; it carries its own directional exposure. The exact leverage level is not clearly stated, so I will not put a number on it. What matters is that the bot can switch from long to short when the votes change.
Bitcoin backtests: strong history and one terrible year
For bitcoin, the full tested history runs from 2017 to October 2026. Starting with 10,000 USDT, buy-and-hold ends near 60,000 USDT, with a maximum drawdown of 83.84%. Strategy 1 ends higher, with a maximum drawdown of 77.38%. In other words, even the trend strategy still suffers a drawdown close to 80%. Most people would have abandoned it well before the recovery.
The worst-case test is a single bear-market year, bought at the start of that year. Buy-and-hold ends near 2,000 USDT with a maximum drawdown of 81.34%. Strategy 1 ends near 4,500 USDT with a maximum drawdown of 63.95%. That is better than the market, but the capital is still more than halved. Large drops still produce losses.
Ether backtests: spectacular numbers, then a wake-up call
For ether, I also looked at a period covering the FTX crisis, from November 2021 to the end of December 2022. Starting with 10,000 USDT, buy-and-hold ends near 2,800 USDT. Strategy 1 ends near 6,000 USDT, down about 39%, with a maximum drawdown of 53.32%. Strategy R ends near 9,000 USDT, still a small loss, with a maximum drawdown of 47.81%. Strategy R may have lost money, but it lost less than the market in that period.
The full ether history since 2017 shows more dramatic gaps. Buy-and-hold ends near 80,000 USDT with a maximum drawdown of 94.07%. Strategy 1 ends near 1.7 million USDT with a maximum drawdown of 66%. Strategy R ends near 7.7 million USDT with a maximum drawdown of 61.87%. Those numbers look spectacular on a screen, but they say nothing about the future.
The 2023-2026 period nobody wants to show
Run the same bitcoin comparison from January 2023 to October 2026, and the story changes. Buy-and-hold ends near 49,000 USDT. Strategy 1 and strategy R both end near 31,300 to 31,400 USDT. Both strategies made money in absolute terms, but they underperformed a simple buy-and-hold. This counter-example matters because most people only highlight the periods that make a system look perfect.
This is why I keep saying that past data is not a guarantee. A strategy can beat the market for years and then lag for several years. The market is alive. Any rule-based system can stop working when conditions change. The numbers are useful for understanding risk, not for predicting your account balance.
How I separate working money from core holdings
If you decide to explore this, I would not tell you to put all your funds on an exchange or in a bot. My own habit is to split money into two parts. The first part is working money, the amount I am willing to leave on an exchange or a bot for experiments. I treat exchange platform risk as a separate line. The second part, and the larger core, is self-custody in a cold wallet. If the keys are not in your own hands, the coins do not really belong to you.
For that self-custody part, what matters most is keeping the private keys under your own control.
My recommended workflow is simple. Find a strategy, have your AI reverse-engineer the rules, gather data, and run backtests. Then build your own bot. It can execute automatically or ask you to confirm each buy and sell. Keep the rules, the funds, and the on/off switch under your own control. If the strategy stops working, find a better one. Never hand over control.
The real prize is control, not a magical bot
AI agents gave me the ability to test public strategies quickly, but they did not remove the two outcomes of trading. The bitcoin and ether backtests show why: a strategy can deliver huge paper gains over the full history, suffer deep drawdowns, and still underperform buy-and-hold from January 2023 to October 2026.
The practical lesson is to stay in control. Use AI to do the heavy lifting: reverse-engineering, checking data, writing backtests, building the bot. Treat it as an employee that follows rules, not as an oracle. If you explore new tools, that is fine. Just be prepared to take responsibility for every decision, and keep the keys, the rules, and the switch in your own hands.