Why Not Equal Weights?
Across 12 walk-forward validation rounds (train on the past, test on a future the model never saw, slide the window forward, repeat) spanning three Bitcoin market cycles, one indicator finished first in 6: Power Law Position. No other feature won more than 3, and none won in more than one test window. The SHAP analysis (SHAP is an itemized receipt for a model's forecast: it scores how much each input moved the output) was not unanimous, but it was lopsided (with caveats about model size that the SHAP analysis documents).
The honest answer to "why not equal weights" turned out to be narrower than this article once claimed. A June 2026 test ran the full DCA backtest (DCA is dollar-cost averaging: buying a fixed amount on a fixed schedule, whatever the price) under three weighting schemes: the published weights, equal weights, and weights re-derived walk-forward with no future knowledge (see the backtest analysis). Equal weights actually produce larger portfolios, because they read conditions as favorable more often and deploy more capital. What the SHAP-informed weights buy is capital efficiency: more value per pocket dollar than equal weights at every tested entry point, because the heavier valuation tilt sells harder at cycle peaks. They are a calibration choice on the aggressiveness-versus-efficiency dial, grounded in the rankings, not a source of timing edge.
The 13 Indicators and Their Weights
Weights are informed by SHAP feature importance (XGBoost + LSTM, walk-forward validated 2014-2026) and then fixed by hand; they are an editorial mapping of the rankings, not a computed output. Power Law at 17% reflects its record as the most frequent #1 feature (6 of 12 rounds, top five in 8). The full reproducible counts, including the indicators whose weights rest on economic rationale rather than SHAP rank, are in the SHAP analysis.
Valuation (41%): Power Law Position (17%), 200-Week MA Distance (12%), MVRV Ratio (7.5%), Puell Multiple (4.7%). The heaviest cluster because SHAP consistently identified long-term valuation as the most important dimension.
On-chain demand (11%): BTC Transaction Fees (11.2%). SHAP #3 overall.
Mean reversion (9%): Mayer Multiple (9.3%). SHAP #4.
Trend/Cycle (20%): Pi Cycle Top (8.4%), 50/200 DMA Cross (6.5%), NUPL (5.6%).
Macro (9%): M2 Money Supply (6.5%), DXY Dollar Index (2.8%).
Momentum + Volatility (8%): RSI (4.7%), 30-Day Volatility (3.7%).
The Formula
Each indicator produces a daily grade (A through F) that converts to points: A = 10, B = 7.5, C = 5, D = 2.5, F = 0. The composite is a weighted average:
Score = sum(gradePoints[i] x weight[i]) / sum(weight[i])
Missing indicators are excluded rather than scored as neutral. The score ranges from 0 to 10, with a historical mean of 5.46.
Verdicts
The score maps to buy/sell verdicts. The boundaries were calibrated by Bayesian optimization, an automated search that tries many boundary settings and keeps the ones that score best (Optuna is the tool, 1,000 trials), against 2014-2021 data, then evaluated on a held-out 2022-2026 window they were never optimized against (years the search never saw, so it could not tune itself to fit them); full methodology in docs/BACKTEST.md:
5.5 and above: Strong Buy (5x your base amount)
4.5 to 5.49: Accumulate (3x)
4.1 to 4.49: Hold (1x)
3.8 to 4.09: Reduce (0.5x)
Below 3.8: Take Profit (sell 2.3% of stack per day)
Cash from Take Profit is redeployed during the next buying period. Market profits fund the next accumulation, not your wallet.
Why the Verdict Flickers, and Why the Flicker Is Noise
There is a design tension worth stating plainly. The Hold band is 0.4 points wide and the Reduce band is 0.3, but the composite's typical day-to-day wobble is larger than either: its daily standard deviation is around one full point. A score sitting near a boundary can cross it and cross back on an ordinary trading day, flipping the verdict twice without anything real having changed. The Markov analysis (a tally of how often each verdict leads to each other one) measured exactly this: about 56% of all historical daily verdict changes reversed within three days. More than half of the verdict's "changes" are threshold noise.
That sits awkwardly next to a dashboard you can open every morning, so the reconciliation should be explicit rather than buried. A daily check reveals almost nothing a weekly check would not. The dashboard now says so under the verdict: it reports how many days the current reading has held, with a note that a one-day flip is usually noise. The churn is surfaced rather than hidden behind a smoothing rule, because any smoothing buffer is an arbitrary free parameter (a knob with no principled setting: why three days and not five?), and an honest instrument should show its own jitter rather than paper over it. Read the persistence, not the flip.
Composite vs Power Law Alone
The backtest analysis showed that Power Law Position alone, the most frequent #1 in the SHAP rankings, produces a better return per dollar invested than the full 13-indicator composite in 9 of 12 tested entry points. The composite builds a larger total portfolio in all 12.
The reason lies in the grade distributions (computed with data through 12 June 2026, from a 2017 start). Power Law calls for "hold at base amount" on 44% of days; the composite calls for "buy 5x" on 55%. Power Law also sells more often, with a D grade firing on 11% of days, generating more cash for redeployment.
The other twelve indicators add context for reading market conditions, and each has its own chart and grade on the dashboard. Averaged into a composite, they do not improve DCA timing. That tracks with the SHAP finding. Long-term valuation is the strongest signal, and Power Law captures it, though the corridor it measures is a fitted trendline rather than a structural law, which caps how much weight the reading can bear.
What the Score Does Not Do
The composite score is not a price prediction. The backtest shows it does not beat Standard DCA on pure timing: same dollars in, same result out. Its value sits in capital allocation: flagging the days when conditions favor deploying more.
The score updates daily and moves slowly. It is built for investors who measure time in months, not minutes.