Methodology

The Illusion of Diversification: Why 13 Indicators Are Really 7 Dimensions

April 2026

TL;DR A correlation matrix across all 13 dashboard indicators shows many are tightly clustered. Two genuinely independent indicators (DXY, Volatility) joined the set. Cluster-corrected weights were tested, then dropped. SHAP's original rankings won the backtest. The valuation cluster carries 41% of weight because that's what the ML found matters most.

The Assumption

Power Law Position and the 200-Week Moving Average correlate at 0.93. Two indicators, one signal. They sit on the dashboard as separate rows, with separate weights, telling the same story in slightly different fonts.

That was not the plan. The thirteen indicators were picked to cover distinct dimensions of the bitcoin market: valuation, momentum, trend, on-chain activity, miner economics, macro conditions. The weights came from SHAP feature importance, derived from machine learning models trained across four complete bitcoin market cycles. Rigorous numbers, derived from data rather than taste.

Then came the correlation matrix. Of the thirteen indicators, only seven move independently. The rest are echoes. This article walks through what the matrix shows, what was changed in response, and what was left alone.

What the Correlation Matrix Shows

The matrix uses Pearson correlation coefficients across 4,322 daily observations, from June 2014 to April 2026. A coefficient above 0.7 means two indicators measure substantially the same thing.

Cluster 1: Valuation (r = 0.71 to 0.93)
Power Law Position, 200-Week MA Distance, MVRV Ratio, and Puell Multiple all cluster tightly. They answer variations of one question about whether bitcoin sits cheap or expensive against its own history. Different math, same underlying signal. The tightest pair is Power Law and the 200-Week MA at r = 0.93.

Cluster 2: Trend and Cycle (r = 0.73 to 0.86)
Pi Cycle Top, 50/200 DMA Cross, and NUPL form a second cluster. All three respond to the direction and maturity of the current trend. Pi Cycle and DMA Cross at r = 0.86 are both moving average relationship indicators.

Bridges: Mayer Multiple sits between clusters (r = 0.87 with Puell, 0.82 with MVRV, but also 0.53 with RSI). BTC Fees is adjacent to valuation (r = 0.50-0.65) but captures a distinct signal: on-chain demand for block space.

Independent: RSI (max r = 0.53), M2 Money Supply (max r = 0.46), DXY Dollar Index (max r = 0.46), and 30-Day Volatility (max r = 0.45) are all genuinely independent. No pair among them or with the clusters exceeds 0.50.

What Changed

Three steps followed.

1. Added independent indicators. A shortlist of eight candidates from the ML pipeline was screened, and two passed: DXY (Dollar Index) and 30-Day Realized Volatility. Both clear the independence bar with max |r| < 0.50 against every existing indicator. DXY contributes a macro and currency dimension. Volatility captures regime shifts and capitulation. The count of genuinely independent dimensions rose from two (RSI, M2) to four.

2. Cluster-corrected the weights, then reversed course. The first attempt cut valuation from 44% to 30%. Backtesting punished the change. SHAP weights, with Power Law at 18%, produced sharper sell signals because the composite drops further at cycle peaks. The dashboard reverted to SHAP-derived weights, with DXY (3%) and Volatility (4%) added as small independent contributors and the whole set renormalized.

3. Discovered that Power Law alone outperforms the composite for DCA timing. A single-indicator strategy using only Power Law Position, with its D and F grades triggering sells, produced the best return per dollar of capital deployed. The other twelve indicators add context for reading market conditions, but they do not improve DCA outcomes when averaged into a composite. The full three-way comparison sits in the backtest analysis.

What Stays

All correlated indicators remain on the dashboard. Each earns its keep for individual chart exploration:

  • Power Law shows where price sits in the structural growth corridor
  • 200-Week MA shows distance from the most reliable support level
  • MVRV shows on-chain cost basis: different data source than price-derived indicators
  • Puell shows miner economics: the supply side

They correlate because they respond to the same market cycles. But each tells a different story about why. A reader exploring the charts gets value from seeing all four angles on the same moment. The trouble was never displaying them side by side. It was adding their scores.

Remaining Limitations

  • The SHAP weights give 41% to the valuation cluster. This is intentional, since backtesting showed it outperforms a balanced allocation, but it means the composite is heavily driven by one dimension.
  • The composite reflects bitcoin's history as a structurally rising asset. The weights and thresholds have not been tested against a sustained multi-year drawdown or a flat decade.
  • Power Law alone outperforms the composite for capital efficiency in backtesting. The twelve other indicators add context but do not improve DCA outcomes. See the backtest analysis.
  • The full correlation matrix is published in docs/CORRELATION.md in the project repository for independent verification.