Show HN: TrendFi – I built AI trading signals that self-optimize
For the past few years, I've been obsessed with trying to get AI to produce accurate trading signals. The main challenges I found with AI trading models are lack of consistency, context window bottlenecks, hard to backtest, and high cost.
Asking ChatGPT "Should I buy Bitcoin today?" doesn't work well because the LLM doesnt have a set trading strategy to opperate from. In addition, the small context window makes it challenging to fit enough historical data into. Not to mention it gets very expensive as well.
My solution is a hybrid approach. Instead of having the LLM make direct predictions, I use AI as the "conductor" in the system that has the ability to run a series of backtesting simulations on high thread count AMD EPYC servers.
The AI then processes the results, identifies optimal parameter changes, and can self-optimize after every trade (as needed) to adapt to changing market conditions. This setup uses LLMs for high-level reasoning and CPU cores for what they do best.
I should mention that TrendFi is designed as a trend based model. It's not for day trading or catching every small market fluctuation. Its goal is to accurately identify major trend changes and provide clear alerts for those key moments.
Check it out: https://trend.fi
– Michael
If you exclude all the crypto related stuff (a lot of people here don’t want to touch that) the 2024 average individual trade return was a bit over 7%, strongly boosted by 3 outlier trades. Hard to calculate the fund performance for that subset though. Meanwhile if you bought and held MAGS over the same time period you would have gotten about a 1.67x return with long term cap gains tax benefit - probably a better total play than this model even with the crypto trades (especially if accounting for crypto trade fees).
Also there is no indication of YTD 2025 performance, which would be more indicative of model strength due to the March-April dip and slowdown in crypto + AI bull markets. I’ve been burned before by services like this and am not in a hurry to do so again.
If your performance is correct and you're getting 200% per year, why sell the sauce for $20 a month?
I would just post the historical and open trades. Charge a subscription if you want to get exactly when you're selling them, otherwise release them on a few week lag to incentivize subscription. But give the rest away for free as you want others to pile in to your trade.
The best model I tried was from The Technical Traders. They offer a trend based model for SPX/QQQ and they consistently outperform buy and hold for 10+ years now.
We were inspired by their service and wanted to take a modern spin on trend based models with a more affordable price point (they charge $2500yr).
Trend changes don't happen all that frequently. Therefore, they wont disappear immediately either. Timing trend changes correctly can make a huge difference.
- Has that been independently verified?
- Have they only been trading for 10 years?
- How do we make sure this isn't a case of chance + survivorship bias?
- What does "they're the real deal" mean?
- Two years is not enough to test a strategy like this. 20 years would be better better, but it sounds like they already did that and the strategy fails half the time (hence the 10-year claim)? Does it fail randomly, or did they market a strategy that worked for the first 10 years, then didn't, and now have a new strategy with the same claims?
I'm asking these questions because people usually work hard for their money, and when someone makes (or relays) a claim that they can consistently multiply it better than expected, skepticism and scrutiny is imperative.
Because TrendFi is a trend based model, the signals are not for HFT or day trading. Instead, we identify major market shifts that play out over weeks or months. This long timeframe means the signals are less vulnerable to the immediate front running or spread manipulation that plagues short-term signals.
Also, the AI's ability to self optimize after trades means the underlying parameters are not static. If external players were to try and consistently game the signals, the system is designed to adapt to those new dynamics.
Trend based models are different. The "secret" isn't a secret at all. Think of a trend based trading strategy more as a disciplined methodology. Trend based models follow price and attempt to identify as early as possible when an asset flips from bearish to bullish.
Therefore, selling our model wont destroy the underlying trend, especially in highly liquid assets, such as: SPX, BTC, etc.
A potential argument is that these signals are only applicable to a certain bracket of portfolio sizes (e.g. larger AUM funds would not be able to trade this strategy) -- but you are sharing this with folks presumably in your range of portfolio size.
This is why Wall Street is able to hire talented young traders, some of whom will develop profitable systems. Over time, some of them are able to amass enough savings and the track record necessary to get investors. But at t=0, somebody has to pay the rent.
The person buying the secret will pay less than the long term profit as a fee for them giving you the money upfront now.
But if you have rights to resell the secret multiple times you will earn money far quicker, perhaps more than the secret’s long term value.
In other words, markets have varying scales of information and are certaintly not efficient despite what academia has you believe. Trend following works precisely because there can be underlying reasons for the trend, and traders piling on just reinforces that trend. Where profitability of the signal depends finding robust trend changes as early as possible.