Signals and Features¶

This page covers model-side feature engineering and signal composition before framework allocation. In the hardened runtime, strategies consume provider-backed, as-of-materialized, observed-only feature frames.

Provider-backed feature flow¶

The framework now owns raw feature sourcing.

Strategies declare required_feature_sets().
Registered feature providers build lazy, as-of-filtered feature pipelines and the registry collects once after joining the observed feature set.
Strategy hooks still receive an eager ctx.features_df only for start_date..current_date.
Strategy methods must not load external files or make network/database calls directly.
The weight path now keeps date alignment, preference/profile preparation, and calendar normalization in Polars. NumPy is reserved for the final sequential allocation kernels where order-dependent state still matters.

Primary signals¶

Signal	Weight	Meaning
MVRV value	70%	Lower MVRV increases buy intensity
Price vs MA	20%	Below 200-day MA increases buy intensity
4-year percentile	10%	Cycle-context signal
---	---	---

Key Concepts for Analysts¶

If you are new to Bitcoin analysis, here are the primary metrics used in this model:

MVRV (Market Value to Realized Value): A ratio that compares Bitcoin's market capitalization (current price × total supply) to its realized capitalization (the value of all coins at the price they last moved).
- High MVRV suggests the market is in high profit (potential overvaluation).
- Low MVRV suggests the market is in loss or near "cost basis" (potential undervaluation).
Price vs 200-Day Moving Average: A standard trend-following metric. Prices below the 200-day MA often indicate long-term "value" zones for accumulation.
Sats (Satoshis): The smallest unit of Bitcoin (1 BTC = 100,000,000 sats). "Stacking sats" is a community term for regular, small-scale Bitcoin accumulation.

Feature construction¶

200-day moving average¶

price_ma = price.rolling(200, min_periods=100).mean()
price_vs_ma = (price / price_ma) - 1

MVRV Z-score and gradient¶

mvrv_zscore = (mvrv - rolling_mean(365)) / rolling_std(365)
mvrv_gradient = tanh(mvrv_zscore.diff(30).ewm(span=30).mean() * 2)

4-year percentile and volatility¶

mvrv_percentile = rolling_percentile(mvrv, window=1461)
mvrv_volatility = percentile_rank(mvrv_zscore.rolling(90).std())

Leakage prevention¶

Signal columns are shifted by one day so weight for day t uses only information available through t-1.

signal_cols = [
    "price_vs_ma",
    "mvrv_zscore",
    "mvrv_gradient",
    "mvrv_percentile",
    "mvrv_acceleration",
    "mvrv_zone",
    "mvrv_volatility",
]
features[signal_cols] = features[signal_cols].shift(1)

Strict validation also rematerializes features from truncated and perturbed source data, runs purged walk-forward checks, block-shuffled null tests, block bootstrap confidence intervals, and feature-drift diagnostics.

Dynamic multiplier structure¶

combined = value_signal * 0.70 + ma_signal * 0.20 + pct_signal * 0.10
combined *= acceleration_modifier
combined *= confidence_boost
combined *= volatility_dampening
dynamic = exp(clip(combined * DYNAMIC_STRENGTH, -5, 100))