# Treasury curve → BTC event study — 2026-08-20

## Verdict

Do not add a 30-year-yield feature to Pine or the MLP yet. The first point-in-time-safe screen does not support a reliable short-horizon long-BTC signal, and it has no usable current OOS events.

## Method

- BTC: the local daily Coinbase BTCUSD TradingView export, 2014-12-01 through 2026-06-25.
- Rates: FRED daily constant-maturity 2Y (`DGS2`), 10Y (`DGS10`), and 30Y (`DGS30`) yields.
- Availability: every Treasury observation is shifted by one BTC day before it is used. This prevents an end-of-day yield from explaining same-day BTC.
- Fixed hypotheses, assessed at 1, 5, 10, 20, and 60 BTC-day horizons:
  1. `DGS10 > DGS30` (10Y above 30Y).
  2. 2Y falls at least 25 bp over 10 days.
  3. All of 2Y, 10Y, and 30Y fall at least 25 bp over 10 days.

## IS evidence through 2026-02-28

| Condition | Sample days | 5D median return | 10D median return | 20D median return | Baseline 5D / 10D / 20D median |
|---|---:|---:|---:|---:|---:|
| 10Y > 30Y | 45 | -0.15% | -0.47% | +2.48% | +0.59% / +1.15% / +1.86% |
| 2Y sharply falling | 104 | -0.05% | +0.14% | +3.20% | +0.59% / +1.15% / +1.86% |
| All three sharply falling | 47 | -0.20% | -2.84% | +1.86% | +0.59% / +1.15% / +1.86% |

The visual pattern is plausible in selected episodes, but the aggregate data does not show the proposed immediate upside effect. In particular, broad yield declines were worse than baseline at five and ten days. The 10Y-above-30Y result has only 45 observations and a negative 60-day median return (-10.47%), so it is not a stable regime signal either.

## OOS reality check

The available OOS segment is 2026-03-01 through 2026-06-25. It contains zero occurrences for all three fixed conditions, so it cannot validate or reject them. This is a reason not to promote the feature—not evidence that it works.

## Recommended next experiment

Leave the live strategy unchanged. Keep collecting daily data and rerun the same fixed report. If future OOS produces a meaningful event count and the same direction survives, test one derived 10s30s feature in a separate feature-ablation run. Do not add raw 30Y, alter exports, or spend Pine security budget before that gate passes.

## Reproduce

```bash
python tools/run_treasury_curve_event_study.py --download-fred
```