Topic: Help Needed: Calibrating My AI (Read 273 times)

I know it can’t be completely flawless, but just check out how spot-on some of these forecasts are:
https://x.com/lets_trade_ai/status/1857940039545610614
https://x.com/lets_trade_ai/status/1856395610913472885

Indicators are algorithms. But a machine learning model operates in a completely different way.
My AI is like playing chess with a chart, but it only calculates the opponent’s moves.


It definitely works, but only when combined with experience. That’s why I’m asking if there’s someone who trades professionally and can manage risks based on my data.

 Indicators for me isn't really accurate and it's lagging and it seems basing only on the previous results. So if you want to use indicators i suggest use it for confirmation only and must base on price action as its more accurate than that especially if there's a clear trend, for sure it will give you a good results afterwards.

Yes, it's true that a lot of technology is currently advancing through AI. But But it is not possible to do 100% correct analysis through AI. I know, trading requires a lot of experience because without experience success is not possible and observation is not possible to analyze everything in the market, but those who do not have experience can take help from another training expert. I have also been learning for the past year, I will end up learning something.

Did I understand correctly that you have a lot of trading experience?
Maybe you could help me with a strategy, please?

I’ve calibrated the AI, and the results are exceeding expectations.

Approximately 150 predictions for each of the 18 cryptocurrencies over the past month. Half of the coins fit into ideal statistics due to non-volatile predictions of 2%-5%, which were easily fulfilled over time.

Right now, I'm experimenting with different coins and just keeping those that show good statistics, while removing the others. Tracking down the reason why some assets are predictable while others aren’t remains a mystery to me.

Here’s how a good forecast and an anomalous one look:
https://i.ibb.co/Vt4X0bw/photo-2024-11-11-14-43-47.jpg
https://i.ibb.co/19m71jS/photo-2024-11-11-14-44-48.jpg

Yes, it’s possible to encrypt the code, but I prefer a security approach based on the black box principle.

That's interesting. How many simulations have you done? What's the biggest factor that determines this success rate? 90% sounds too high even if your model is really good. It's just hard to believe based on my experience.

Are you afraid that the VPS/host will record your source code? Can't you package or encrypt it before running it on other computers?

The public test itself has already been launched.
I’ve also added an algorithm to check how effective the predictions are. For BTC, the success rate is currently around 75%.
Some coins have a 0% success rate, while others are at 90-99%.
Each coin needs to be calibrated individually.

As for a GPU VPS server — it could definitely make things easier, but I’d really prefer not to publish the code on external platforms.
The issue isn’t even the code itself, since all the ML models are public, but rather the specific AI configuration.

That's a beefy PC alright. Are you saying that you'll host a demo or public test on your rig with this spec, or are you looking for a VPS/server that can grant you this spec for hosting the code?

Have you looked at the renting GPU platform that has been popping out lately? If I remember correctly some of them offer less than $1/hr for renting a single 4090, not sure about the rest of the specs though. You can probably order a custom plan, although I'm not sure they'll agree if you only require one machine. CMIIW.

I’ve tested around 30 different indicators and calibrated them for various timeframes and the lifespan of each instrument. The results were mixed: some coins always showed sideways movement, while others, like STRK, predicted a +20% gain just before a pump. In fact, when 90% of the coins indicate a long, a long actually happens.
https://i.ibb.co/m8cfSbL/photo-2024-10-21-02-22-42.jpg
https://i.ibb.co/LvvSpLz/photo-2024-10-21-02-18-00.jpg

However, the issue is that the AI struggles to accurately pinpoint the start of local corrections. It feels like something is missing...


It's indeed AI, if we can even say that ML equals AI.

The logic of your AI is correct, but that's not how it works. A simple language model can't function the way you describe; everything needs to be built modularly.
First, you need to calibrate the pattern recognition in regular price fluctuations. Sentiment and news analysis is the second module, but that requires more power, like Amazon or Google Colab.

There are also modules 3 and 4, but that's another story.


I need a PC (2x4090, 128GB RAM) that can process the maximum dataset in under one minute. Right now, it takes about an hour.
The AI is already operational, but it's currently working with a dataset that has limited price history detail.

Here's the link: https://t.me/ai_lets_trade_1h

Every hour, it gathers new market data, trains on it, and sends a report to the channel. However, the report is incomplete—it only includes the minimum and maximum forecasted prices.
In reality, the forecast represents a continuous stream of prices. The module that will display this is still under development, as is an additional AI that will analyze the forecasts generated by this AI.

What does AI mean in this case? I'm afraid it's a buzzword to gain attention.
Is your AI forecaster really an intelligence? Can it check Elon Musk's twitter, read his posts, analyze them (if he posts about Doge going to the moon) and make a prediction? I want to know what's AI in your software, is it capable to think on its own?
I think that you should train your AI to check different indicators, make a prediction on each coin with each indicator and then it will collect the data about which indicator works best for which coin. If that's truly an AI, then it should be able to achieve that level itself.

Yeah, it would be also wise to test it all on one coin for the time being, before expanding into different branches of the market and the coins on it. All do have their quirks indeed.

I think combining a mix of trend-following indicators with momentum-based ones can help balance signals, especially across different timeframes. For instance, Bitcoin often responds well to long-term EMAs, while some altcoins might benefit more from oscillators on shorter timeframes. However, I've noticed that different coins react differently to these indicators.

Combining indicators can indeed be tricky. Many traders find success using a blend of MAs, RSI, and MACD for different coins. Each coin has its own quirks, so you might need to fine-tune per asset. Consider backtesting with historical data to find the best combinations.

A demo, which joniboini suggested, would be a good way out too.

what resources do you need?

The AI already uses basic indicators, but I want to figure out how to combine them. Or maybe use all of them at once, it's still unclear.

I'm open to a public test and am trying to launch a beta version, but I don't have enough resources yet.

Thanks for the idea about open-source trading bots, I hadn’t thought of that before.

If you know about trading, you should know exactly what indicators are and te purpose they serve to general technical analysis.
Indicators to the best of my knowledge are not special to any coins and were not created specifically for any coin so there is no particular indicator that works best for a certain coin.
If you built an AI without really understanding how trading works, how can you expect success.

Why not make a demo that allows people to backtest their strategy so you get the data to calibrate your AI? You need to be explicit about this so users don't feel tricked. Have you looked at open-source trading bots and checked their default algorithm as a standard to develop your own AI trading strategy? I think you'll have to run multiple tests eventually so your AI can recognize more patterns, assuming you have the resources to do so. CMIIW.

I don't think you can find indicators that work best since it's all depends in your setup and strategy.

If I were you, why not test every indicator available? You can start with the major indicators like RSI, MACD, Moving average, and Bollinger Band.

If you don't know how these indicators work, I suggest watch a video from Trade Prime for a better explanation of these indicators and which the best combinations is the best.

Here's the link: https://www.youtube.com/@TradePrime/videos

There is no one specific indicators that can be 100% accurate in the market. Heck, even combinations of indicators will not get you 100% positive results. However, you can tinker around different indicators to produce a good win rate with a nice risk reward ratio.

Perhaps you can check out the most popular community scripts on trading view and modify them if you want (click on indicators). If there's any script or trading strategy that is available, I am pretty sure it's not readily available to the public.

Hey everyone!

I’ve built an AI that forecasts prices for 18 different coins—Bitcoin included. But I could really use some help fine-tuning it, especially when it comes to technical indicators.

If you’ve noticed which indicators work best for certain coins, or spotted any unique patterns on specific timeframes, I’d love to hear your thoughts. Maybe you’ve got insights I’m missing?

I’d be grateful for any help you can offer. I might code like a wizard, but my trading skills... well, they could definitely use some backup 😅.

Thanks a ton in advance!

DM_tg: letstrade_ai