Neural networks are a class of machine learning models loosely inspired by how the brain processes information. In crypto analysis, they are applied to price data, volume, on-chain metrics, and social signals to detect statistical patterns that repeat across market conditions.

The basic structure

A neural network consists of layers: an input layer that receives raw data, one or more hidden layers that extract features, and an output layer that produces a result — in our case, a composite score between -1 and +1.

Each connection between neurons has a weight. During training, the model adjusts these weights using backpropagation to minimise prediction error on historical data. After enough training steps, the weights encode the patterns the model has learned.

What goes in

The Coinblockers model receives normalised inputs across four domains:

  • Technical: 14-period RSI, MACD histogram, price vs 20-day and 50-day moving averages, 7-day vs 30-day volume ratio
  • On-chain: exchange net flow, active addresses (normalised), miner revenue trend, NVT signal
  • Social: weighted feed of platform mentions and sentiment polarity
  • Market structure: 30-day volatility, correlation with Bitcoin, liquidity score

Each input is z-scored against a 90-day rolling window before it enters the network. This prevents large-magnitude inputs (like raw volume in billions) from dominating smaller but meaningful signals.

What the output means

The model outputs a number. If that number exceeds a threshold, the rule-based engine considers opening a position. The threshold is calibrated to historical false-positive rates — we aim for precision over recall, meaning fewer signals but higher confidence per signal.

The composite score displayed on each coin's page is this output, mapped to a 0–100 scale for readability.

What neural networks cannot do

  • They cannot predict the future. They identify patterns that repeated in historical data, with no guarantee those patterns continue.
  • They are opaque. A neural network cannot tell you why a score is high in plain language — only which inputs contributed most (via SHAP values or similar attribution methods).
  • They overfit. A model trained on 2020–2023 data may perform poorly during a regulatory shock or macro event outside its training distribution.
  • They amplify what humans put in. Garbage inputs produce garbage outputs.

Why we publish the methodology

Every factor, weight, and threshold in the Coinblockers scoring model is published on the Methodology page. This is intentional. A model you cannot audit is not useful for education. We want you to understand exactly what the score means before acting on any signal — and to understand that this is a hypothetical educational tool, not financial advice.

Neural networks are one tool among many. When they agree with your own analysis, that convergence might be worth noting. When they disagree, understanding why is often more valuable than the signal itself.