Preliminary

2026, actually applied

What each validated piece would have returned since 1 January 2026 — including the uncomfortable result.

The question: Outside the tests: in one concrete, recent year, what would have happened?

What was found

A fourteen-year backtest is easy to view from a distance. A year in progress isn't. Over 249 days of 2026, with 10 basis points of costs and the drawdown model trained ONLY on pre-2026 data plus a 30-day purge, the uncomfortable result goes first: the overlay that was live in production made 2026 far worse than doing nothing. It runs between 1.0x and 1.8x, averaged 1.50x all year because BTC spent 2026 below its trend, and in a down year that multiplies the loss: BTC's −9.4% becomes −24.1%. Publishing this is the reason the rest of this section can be read with confidence.

-64.7%-44.8%-24.9%-5.0%14.9%-9.4%-39.7%Buy and hold BTC-24.1%-57.0%Power-law overlay (retired)-10.1%-34.9%Overlay unlevered1.4%-26.0%Overlay + risk control7.2%-35.9%Crowding book2026 returnMaximum drawdown
2026 through 6 September, 249 days, 10 basis points of costs. Return and maximum drawdown for each portfolio.analysis/scripts/phase7_ytd2026_backtest_2026-09-07.py

Try it yourself

The test any concentrated improvement deserves: how many days does it come from? Drag the two numbers and compare with the whole year.

Total improvement (percentage points) = Days x Return avoided on each
Improvement those days contribute

What share of the year they are, out of 249 days:

The default values give the real +20.8 points risk control contributed in 2026, out of 6% of the year. That is the number to look at before the return.

Illustrative example numbers for practice — not real data.

How it was tested, step by step

  1. Each validated piece is taken and left to run over 2026 without any refitting, with 10 basis points of costs.
  2. The drawdown model is trained only on pre-2026 data, with a 30-day purge, so no 2026 information reaches the 2026 signal.
  3. A version of the overlay normalized to a 1.0x maximum is included — what somebody with only spot can do — to separate the leverage effect from the shape effect.
What this does NOT say. And the good result shouldn't be believed either. Risk control contributes +20.8 percentage points, but they come from 16 days out of 249 — 6% of the year — in which BTC averaged −1.38% a day against −0.01% for the rest. An improvement that concentrated over 249 days is the size of luck, not of skill. The crowding book warns in the other direction: it was up 40% in late August and gave back to +7.2% in nine days, with −10.3% in a single session. Market-neutral does not mean low-risk.

Tracking

This finding has no live number: it was measured on archived data, and putting a chart here would imply continuous tracking that doesn't exist. The figures are above, with the script that produced them.

← All research