PowerTrader_AI: a spot crypto bot built on a nearest neighbour price predictor and tiered DCA

Four Python files and a short dependency list

PowerTrader_AI is described as a fully automated crypto trading system powered by a custom price prediction AI and a structured or tiered dollar cost averaging system. The repository is small enough to read in an afternoon: the top level holds a LICENSE, the README, four Python modules named pt_hub.py, pt_thinker.py, pt_trader.py and pt_trainer.py, and a requirements.txt.

The four names describe the pipeline. The trainer is the part that builds the memory of patterns. The thinker is the part the README describes as sending the signal to start a trade. The trader is the part that places orders. The hub is the part that deals with the outside world, and given that the dependency list includes requests and psutil, it is the piece most likely to make network calls.

The dependencies are correspondingly modest: requests, psutil, matplotlib, colorama, cryptography, PyNaCl and kucoin-python. The presence of kucoin-python tells you the exchange side is KuCoin. The presence of cryptography and PyNaCl tells you there is some signing or key handling involved, and psutil suggests process or resource monitoring, which is not unusual for a bot that runs unattended. matplotlib is for the charts, and colorama for terminal output.

The licence is Apache-2.0, the last push to the repository was on 2026-04-18, and there are no GitHub releases, so the project is distributed as source with no packaged versions to pin against.

The predictor is a weighted nearest neighbour over stored patterns

The README spends a long passage on the fact that the AI here is not an LLM and not a neural network. It quotes a description of the system as an instance-based, kNN or kernel-style predictor with online per-instance reliability weighting, used as a multi-timeframe trading signal, and then explains what that means in plain terms.

Training works like this. When training for a coin, the system goes through the entire history for that coin on multiple timeframes and saves each pattern it sees, along with what happens on the next candle after the pattern. To produce a prediction, it takes a weighted average of the closest matches in memory to the current pattern in time. That weighted average is computed once for each timeframe, from one hour up to one week, and each timeframe gets its own predicted candle. The low and high prices from those predicted candles are what appear as the blue and orange horizontal lines on the price charts.

The learning happens after the fact. Once a candle closes, the system checks what happened against what it predicted and adjusts the weight for each memory pattern that was used to generate the weighted average, depending on how accurate that pattern turned out to be.

So the model is a memory of historical shapes with a per-pattern trust score that is revised as evidence arrives. The author is upfront that it is extremely simple and still counts as AI, and says the strategy was developed over almost a decade with the AI design worked on for the last few years, based on what actually works in real trading experience rather than on what the author read in LLM responses.

Entry, DCA ladder and the trailing exit

Three separate decisions make up the trading logic, and each is specified precisely in the README.

Entry is driven by the predictor. The thinker's script sends a signal to start a trade for a coin if the ask price drops below at least three of the AI's predicted low prices for that coin, where the prediction covers the currently active candle's high and low for each timeframe from one hour to one week. The three-line threshold is what stops a single noisy timeframe from triggering an entry.

The DCA decision uses either the current price level from the AI, which is tied to how many DCA buys have already been made on that trade, or a hardcoded drawdown percentage for the current level, whichever it hits first. The README gives the mechanics: right after a trade starts, when three blue lines have been crossed, the first DCA does not happen until the price crosses the fourth line, and so on. On top of that there is a hard cap of two DCAs within a rolling 24 hour window, which exists to keep the bot from dumping all the money in too quickly on a coin in an extended downtrend.

Exit uses a trailing profit margin to maximise the gain. The margin is set at 5 percent if no DCA has happened on the trade, and 2.5 percent if any DCA has happened, with a trailing margin gap of 0.5 percent, which is the amount the price has to exceed the margin before the margin starts being raised, so that the profit is maximised once the price falls back below it and the trade is sold.

The author notes that other risk management features could be added with a bit of Python, which is a candid way of saying the risk controls are minimal by choice rather than missing.

No stop loss, and why the author thinks that is correct

The absence of a stop loss is stated up front and defended at length, and understanding that reasoning is more useful than the fact itself.

The author writes that he knows there are commonly essential trading features missing, no stop loss for example, and that this is by design because many of them would not work with the strategy as it stands. The reasons given are a set of personal convictions: he does not believe in selling worthwhile coins at a loss, he does believe in crypto, and he would rather wait and add more money to the account so the bot can buy more of the coin while the price is down. He argues that stop loss and similar tactics are a trick, that adding money to afford more DCA or waiting for extended periods is acceptable, and that anything else is greedy and desperate, the opposite of what is needed for long term growth.

The technical argument underneath is about market type rather than philosophy. This is spot trading, where there is no worry of liquidation, and he argues that many risk management tactics are really only meant for futures trading, and that people apply them to spot trading where they are not necessary.

Whether or not you accept that argument, it defines the risk profile of the system clearly. A position is held until the trailing margin is reached, and the only mechanism for improving the average entry is adding more capital at a lower price, at most twice a day.

Setup, the hub component and the safety caveats

The README carries a setup walkthrough for Windows, and it opens with an unusual admission: the instructions were written by AI and the author asks readers to report errors or issues with the setup process. It says no coding knowledge is needed, that the instructions are Windows based but the software should be able to run on any operating system, and that the page walks through installing PowerTrader AI from start to finish in the order a first time user should do it.

There is a specific warning about existing crypto holdings. If you have any crypto holdings in Robinhood currently, either transfer them out of the Robinhood account or sell them to dollars before going through the setup process, which suggests the setup involves moving funds into the exchange the bot trades on.

The disclaimers are unusually direct. The software places real trades automatically, and you are responsible for everything it does to your money and your account. Keep your API keys private. The author is not giving financial advice, is not responsible for losses, and is not responsible for security breaches to your computer, with the point made that the code is entirely open source and can be confirmed non-malicious. You are fully responsible for doing your own due diligence and for all gains or losses.

The commercial point is stated just as plainly. The author is not selling this trading bot, it is not a product, and the system is for experimentation and education. The only reasons to send money are voluntary donations or buying merch, and the README warns readers not to fall for any scams, insisting the system is free forever. It also notes that forks exist and that the author has not checked them, so cannot confirm or deny their legitimacy.

Where PowerTrader_AI is the wrong choice

The first limitation is that there is no stop loss, and the whole design leans on that absence. A position in an asset that falls by eighty percent is not exited by the system, because the only exit is the trailing margin on a trade the bot opened itself. The DCA ladder helps only if the account can absorb more capital at a lower price, and the two-per-day cap means a long decline accumulates a position slowly while the loss on what is already held grows. A system whose risk control is the user's willingness to add money needs a user with money to add.

The second limitation is the evidence. The author is explicit that the strategy comes from personal trading experience rather than from published results, and there are no benchmarks, no backtest figures and no performance claims anywhere in the documentation. That is honest, and it also means you cannot evaluate the strategy from the README. You can only read the code, run it, and form your own view, and even then the results depend entirely on the market the bot is pointed at.

The third is the operational exposure. A bot that holds API keys and places real orders automatically is a security liability, and the documentation warns about it without minimising the risk. The dependency on a specific exchange through kucoin-python means the bot is tied to one venue's behaviour, and the hub component, which is the piece with requests and psutil available to it, is the part to read first before handing over credentials.

For a team wanting a trading system to evaluate or operate, the realistic alternatives are a well documented exchange SDK with your own strategy written on top, or a reputable third-party bot with published performance and a support contract. Neither is free, but both are legible in a way this project is not. What PowerTrader_AI offers instead is transparency about its own simplicity, which is a real virtue for learning, and a licence that lets you fork it and fix what you disagree with.

Editorial conclusion

PowerTrader_AI fits a developer who wants to read and modify a complete spot trading system with a small, understandable model rather than a black box service, and who accepts that the system holds positions without a stop loss until the trailing margin is reached. It does not fit anyone expecting risk management by default, and it is not a strategy you can evaluate without reading the code, because the author is explicit that the design comes from personal trading experience rather than from evidence. Before running it, read the four modules in the order they execute, remember the author states the setup instructions were written by an AI and asks for corrections, check what the hub component reaches over the network before you give it API keys, and treat the strategy as one position at a time on spot with money you can afford to hold through a drawdown.

Frequently asked questions

Official sources

1. garagesteve1155/PowerTrader_AI on GitHub
2. Issues
3. License: Apache-2.0
4. README