Hybrid AI Crypto Prediction System

This app combines Macro, Whale, and Sentiment AI agents with RSI and ATR technicals. An XGBoost model analyzes it all to spot high-probability crypto breakouts before the crowd moves.

Comment

Inspiration

Being a full-time mom and a career driven software engineer my inspiration is to make money logically through crypto. Last month I invested some amount in crypto and made 10% of the profit this actually triggers me to think in the way that crypto is all about the data and maths and since living the era of AI and prediction models why not to use my own created model to predict what next and if I make 10% profit without any logic then how can i 2x or 3x my profit to play some of the crypto attributes.

What it does

It simply gives the two signals i.e

  1. Bearish (Expecting the market to go down) - Sell (or stay out)
  2. Bullish (Expecting the market to go up) - Buy (or hold)

These signals are derived from six core parameters i.e RSI for momentum, ATR for volatility, individual Macro, Sentiment, and Whale agent scores, and their combined Weighted Consensus Score then a model analyzes these parameters together to classify whether the market setup aligns with a high-probability bullish or bearish outcome

How we built it

This app was engineered as an end-to-end predictive machine learning pipeline deployed on Streamlit Cloud.

1. Multi-Agent Intelligence Engine

We established a weighted multi-agent consensus module to digest market dynamics: $$\text{Composite Score} = 0.40 \times \text{Macro} + 0.30 \times \text{Sentiment} + 0.30 \times \text{Whale}$$

  • Macro Agent (40%): Evaluates macro liquidity trends and overarching market regimes.
  • Sentiment / Retail Agent (30%): Gauges real-time social noise on a scalar from -1.0 (extreme bear) to +1.0 (extreme bull).
  • Whale / On-Chain Agent (30%): Tracks large-scale exchange inflows, outflows, and wallet clustering.

2. Feature Engineering & Quantitative Indicators

We engineer key price-action variables across high-frequency historical bars:

  • Relative Strength Index (RSI~14~): Measures price velocity and overbought/oversold momentum.
  • Average True Range (ATR~14~): Measures dynamic market volatility to scale volatility targets.
  • Exponential Moving Average 20 (EMA_20): Tracks the 20-period short-term price trend using exponential decay weighting.
  • EMA Percentage Distance (EMA_dist): Calculates relative distance (close - EMA_20) / EMA_20 to scale price position independently of asset price magnitude.
  • Volatility-Adjusted Target (target): Defines a binary label (future_close > close + ATR) requiring future price to exceed dynamic market volatility to clear execution noise.

3. Machine Learning Architecture

We formulate breakout prediction as a binary classification task. The target variable $y_t \in {0, 1}$ is defined by a forward-looking hurdle rate over a horizon of $h = 4$ bars: $$y_t = \begin{cases} 1 & \text{if } \text{Price}_{t+4} \ge \text{Price}_t + 1.0 \times \text{ATR}_t \ 0 & \text{otherwise} \end{cases}$$

An XGBoost Classifier is trained on the resulting feature matrix $X$. To prevent structural overfitting on noisy financial data, we enforce conservative hyperparameters:

  • max_depth = 3: Restricts tree depth to prevent memorization of past noise.
  • learning_rate = 0.05: Ensures gradual, stable weight optimization.
  • eval_metric = "logloss": Penalizes overconfident misclassifications heavily.

Challenges we ran into

  1. Eliminating Data leakage

    Eliminating data leakage means making sure our model doesn't "cheat" by peeking into future price data while training. We solved this by strictly separating our current features from future targets and cleaning up missing values so live predictions stay accurate.

  2. Run Multi agent concurrently

    Handle multiagents while keeping in mind that all agents will work by their own prompts and their chosen models. But asyncio make it simpler while gathering all their results and make pass the composite score seamlessly.

  3. Continuous loop of Risk Governor thread

    Since our app doesn't solely depend on media news and financial attributes it also take care on drawdowns of users while calculating their equity and peak.Since this thread runs on every 5 seconds so maintaining the concurrency of the system while updating the specific component only to keep the user experience smooth while using the App is quite challenging.

  4. Model training

    The model isn't just looking at prices; it has to digest a complex mix of quantitative technical indicators (RSI, ATR) and qualitative multi-agent outputs (Macro, Sentiment, Whale scores). Aligning these asynchronous data feeds into a clean, synchronized training matrix without introducing NaN gaps or temporal mismatches is a major architectural challenge.

  5. Deployment on streamlit

    Balancing live charts, model probabilities, and the Risk Governor's circuit-breaker logs in a clean layout required optimizing rendering loops. Ensuring that a triggered risk rejection updates the dashboard instantly without rendering the whole page

    Accomplishments that we're proud of

    We successfully built an end-to-end quantitative trading system that functions as an intelligent crawler, continually ingesting and evaluating market news, whale movements, and financial data. Beyond signal generation, we are particularly proud of our automated Risk Governor, which acts as a protective shield for user capital during volatile market conditions. While cryptocurrency markets are notoriously chaotic, we engineered a resilient, full-spectrum architecture that addresses the critical pillars of modern crypto trading combining multi-agent intelligence, time-series ML, and dynamic risk management into one cohesive pipeline.

What we learned

  1. I learned how to combine real-time qualitative data by multiple agents and feed it into the training model.
  2. I learned how to manage dataframe sanitization by isolating valid feature windows and systematically stripping NaN values before mapping $X$ and $y$
  3. I architected a real-time Risk Governor safety layer acting as an automated circuit breaker to filter predictions against volatility spikes (ATR).
  4. I learned how to manage app state and execution flow in Streamlit

What's next for Hybrid AI Crypto Prediction System

So far, I have worked on single coin prediction on stable coin later I will extend it with more coins and especially target new coins with time-bound event because in most of the cases I saw people make 70x to invested in unstable coin but at right time like Trump coin.

Built With

Share this project:

Updates

Submission history