4. Model Architecture

ZKSeer’s core forecasting capability is powered by a hybrid architecture consisting of three specialized components:

4.1 Primary Flow Model (Temporal Specialist)#

• Architecture: Enhanced sequence model combining convolutional feature extraction with attention-based temporal modeling (inspired by modern time-series architectures such as PatchTST-style and Transformer variants optimized for financial data).
• Training objective: Multi-horizon prediction of next-candle direction, magnitude, and volatility.
• Input features: Normalized OHLCV sequences, returns, volume profiles, liquidity metrics, and token-age embeddings.
• Key design choice: Heavy emphasis on recent context windows while retaining longer-term structural awareness.

4.2 Regime Detection Module#

• Lightweight classifier that identifies the current market regime (trending, ranging, high-volatility expansion, compression, etc.).
• Outputs are used to dynamically adjust forecast horizon, confidence weighting, and guardrails against unrealistic projections (e.g., runaway pump suppression).

4.3 Reasoning & Explanation Layer#

• A fine-tuned language component that translates numerical model outputs into structured natural language.
• Receives the forecast, key supporting features, and regime context to generate clear explanations of the predicted move.

Training Philosophy#

All core models are trained on large volumes of historical public market data with strict temporal separation to avoid leakage. Continuous online refinement is performed using recent market outcomes to maintain adaptability without catastrophic forgetting.

This specialized stack contrasts with pure LLM approaches, which often struggle with precise numerical forecasting and consistent risk parameterization.