Analyzing Core Machine Learning Models Developed for the Invest Flow AI Project

1. Signal Detection via Gradient Boosting and Ensemble Methods

The investflowal.com platform relies on a multi-layered ensemble of gradient boosting machines (LightGBM and XGBoost) to identify short-term price inefficiencies. These models are trained on tick-level order book data, volume-weighted average price shifts, and bid-ask spread anomalies. Feature engineering focuses on entropy of trade intervals and relative liquidity absorption ratios. The ensemble reduces overfitting by blending predictions from trees trained on different temporal slices of market microstructure data. Validation across crypto spot and futures pairs shows a Sharpe ratio improvement of 0.31 over single-model baselines.

To handle non-stationary market regimes, the system retrains base learners every 4 hours using a sliding window of 14 days. This prevents concept drift from degrading signal quality. The final prediction is a weighted soft-vote, where weights are dynamically adjusted based on recent out-of-sample accuracy. This approach filters out noise from flash crashes and low-liquidity periods.

Feature Importance and Interpretability

SHAP values are computed for each ensemble member to rank feature contributions. The top three features consistently are: cumulative delta divergence, order book imbalance slope, and micro-price deviation from VWAP. These metrics directly capture aggressive versus passive trading behavior.

2. Temporal Pattern Recognition with LSTM and Attention

For mid-frequency trend anticipation, the project deploys a bidirectional LSTM augmented with a temporal attention mechanism. Input sequences are 128-step windows of 5-minute OHLCV data, enriched with on-chain metrics like exchange inflow/outflow velocity and active address counts. The attention layer learns to weight recent price action more heavily during volatile sessions and historical patterns during range-bound markets.

The LSTM architecture uses 3 hidden layers with 256 units each, followed by dropout of 0.3 and batch normalization. Training employs a custom loss function that penalizes false positives more heavily in low-volume regimes. The model achieves a directional accuracy of 62.4% on unseen test data spanning 6 months. Inference latency stays under 15ms per symbol, enabling real-time integration.

A secondary variant uses a transformer encoder instead of LSTM, trading off some interpretability for a 4.2% gain in F1-score on high-cap pairs. Both models feed into a meta-learner that decides whether to execute or wait.

3. Risk-Aware Portfolio Allocation via Reinforcement Learning

The allocation module uses a deep Q-network (DQN) with a continuous action space representing position size percentage. The state space includes current portfolio volatility, drawdown from peak, and the current signal confidence from the ensemble and LSTM models. The reward function is a combination of realized Sharpe ratio minus a penalty for maximum drawdown exceeding 8% over a 5-day window.

Training is performed in a custom simulator that replays historical market data with injected slippage and fee models. The DQN agent learns to reduce exposure during regime shifts detected by a hidden Markov model overlay. After 10,000 simulated episodes, the agent consistently outperforms a fixed 50/50 split strategy by 1.8% annualized return while maintaining lower downside volatility.

FAQ:

What data sources do the Invest Flow AI models use?

They use tick-level order book data, 5-minute OHLCV, on-chain metrics (exchange flows, active addresses), and sentiment scores from news feeds.

How often are the models retrained?

The ensemble models retrain every 4 hours. The LSTM retrains daily. The RL agent retrains weekly using a replay buffer of the latest 30 days.

Can the models handle extreme volatility like flash crashes?

Yes. The ensemble uses dynamic weighting to downweight noisy periods. The RL agent has a penalty for drawdowns over 8%, causing it to reduce positions during such events.

What is the typical inference speed?

Ensemble inference is under 5ms, LSTM under 15ms, and the RL agent decision takes under 2ms. Total pipeline latency is below 30ms per symbol.

Are the models interpretable for compliance?

Yes. SHAP values are logged for every ensemble prediction. Attention weights from the LSTM show which time steps influenced the decision.

Reviews

Marcus K.

I was skeptical about ML for crypto, but the ensemble model caught a divergence I missed. Saved me 12% in a week.

Elena R.

The RL agent cut my drawdown by half compared to my manual strategy. The retraining schedule keeps it adaptive.

David L.

Love the SHAP explanations. I can finally understand why a signal fires. The LSTM attention layer is a game changer for timing.