# HV-RSI — Short-Term Mean Reversion — System Test

A survivorship-bias-free evaluation of Glenn Osborne's HV-RSI short-term mean-reversion
system on the S&P 500, on split- and dividend-adjusted prices. The full write-up
— the cash-matched random-portfolio skill test, the in/out-of-sample work, the
capital-efficiency levers, metrics-by-lookback, the sizing out-of-sample test, exposure
analysis, and the annotated sample chart — is the
[**Deep Dive (`index.html`)**](index.html), which is the authoritative version.
This file is the headline summary.

## What it is

Buy a strong stock (above its 150-day SMA) that closes at fresh short-term lows two days
running, via a limit order 3% below the close (filled only if the next day's low reaches
it); hold up to 20 concurrent slots at 10% each; exit on a close above the prior day's
high (the bounce) or a 5-trading-day time stop.

## Setup

- **Universe:** point-in-time, survivorship-bias-free S&P 500 membership (~959 ever-members).
- **Window:** 2005-01-03 → 2026-05-15 (21.4 years). Starting capital $100,000.
- **Prices:** split- and dividend-adjusted (ratio reconstruction from adjusted close)
  before any signal, fill price, or P&L.
- **Disclosure / OOS pivot:** specified by Glenn Osborne January 2025 (author states 2021
  conception); conservative OOS pivot 2025. The 2005–17 / 2018–26 split is a
  temporal-stability check, separate from the disclosure pivot.
- Commissions, slippage, and interest on idle cash are not modeled.

## Headline results

| Metric | HV-RSI system | Buy & hold SPY |
|---|---|---|
| CAGR (raw, 0% on cash) | 4.02% | 10.88% |
| Exposure-adjusted return (in-market only) | ~11.0% | 10.88% |
| Average capital deployed | ~7.7% | 100% |
| Time in market (days holding ≥1 position) | 37.8% | 100% |
| Max drawdown | −13.35% | −55.19% |
| Win rate / profit factor | 65.0% / 1.58 | — |
| N trades / avg hold | 1,096 / 3.9 trading days | — |
| Final equity ($100k) | $232,089 | $908,276 |

## Findings

- **Capital-light by design.** The book holds a position on ~38% of trading days and
  deploys ~8% of capital on average. Raw CAGR is 4.0%, but measured over time-in-market
  only the deployed slice compounds at ~11%/yr, at a −13.4% maximum drawdown against the
  index's −55%.
- **The signal separates from random.** Cash-matched monkeys (identical machinery, random
  selection) cluster near zero on risk-normalized CAR25 (median ≈ 0); HV-RSI clears 100% of
  the 40 seeds in both windows, with each monkey deploying 4–8× more capital at a lower win
  rate (~57–58% vs 65%).
- **It holds up out-of-sample on large caps.** S&P 500 win rate flat (64.9% → 64.9%) and
  profit factor improves (1.55 → 1.61) from the 2005–17 window to 2018–26; CAR25 +4.16% →
  +6.39%. A small-cap (Russell 2000) cross-check on adjusted data is a planned next test.
- **Selectivity is the source of the edge.** No tested lever that deploys more capital
  (looser limit, fewer consecutive-low days, more slots) raises out-of-sample CAR25 above
  the baseline; loosening admits random-like fills that return near-zero CAR25.
- **Sizing out-of-sample.** The Bandy safe-f fraction pins at the leverage ceiling in both
  windows with drawdown-95 well inside the −20% target — sizing is not the binding
  constraint; capital deployment is.

## Reproduce

```
python prototype.py --index SPY_SandP_500 --start 2005-01-01 --end 2026-05-15
python run_oos.py --index SPY_SandP_500
python develop_eval.py --monkey-seeds 40
python car25_eval.py
python refinements.py
```

The data loader and the survivorship-bias-free dataset are internal to the research
platform and not redistributed; the engine and analysis code are complete in `code/`.

Results are historical simulations, not investment advice.