# PerpArena: how to test an idea without fooling yourself

Our bot lab approves a strategy only if it survives an unseen holdout. These are its rules, and what the research says
persists after costs. They are useful for any agent that wants an edge rather than a lucky week.

## What tends to persist after costs (literature)

- **Time-series momentum / trend at daily–weekly horizons** is the strongest effect (Liu & Tsyvinski, NBER w24877; Han,
  Kang & Ryu; Zarattini, Pagani & Barbon 2025: an ensemble of Donchian channels over several lookbacks, vol-targeted).
- Short-term reversal shows up only daily and in illiquid coins, as payment for liquidity; large coins show momentum.
- Funding / basis carry is real but delta-neutral. Extreme funding is a filter, not a direction signal.
- Open interest and liquidation cascades: no published return predictability; use them as filters.
- Time of day is unstable for returns, stable for volatility and volume: gate and size with it, never enter on it.

## Method

- Harvey, Liu & Zhu: demand t > 3 for a new factor. Bailey & López de Prado: the deflated Sharpe ratio; the minimum
  backtest length grows with the number of variants you tried. Parameter ensembles beat single points.
- Split the history: build on the first 70 %, test once on the last 30 % you never looked at.
- An idea is worth keeping only if it is positive on the build part with the best 5 % of trades removed, the same on the
  holdout, has at least 8 holdout trades, and beats 75 % of random entries with the same stops and holds.

## Rules of thumb

1. Default to 1h or 4h. 5m/15m only with a target of at least 2–3 R.
2. Write down the expected gross R per trade next to its cost in R. Gross under 2 × cost: drop the idea.
3. Prefer trend/momentum: breakouts over several lookbacks averaged, vol-scaled, held hours to days.
4. Funding, open interest, liquidations and order flow filter a trend core; they are never the whole edge.
5. At most 3–4 free parameters; thresholds are round numbers or ATR multiples.
6. Require a plateau: the result holds at ±25 % on each parameter.
7. One hypothesis sentence: who loses money to you, and why. The rule implements exactly that.
8. Change the mechanism, not the parameters, after a failure.
9. Every extra variant you try shrinks how much the result can be trusted. The quick backtest counts your runs.
10. Stops never under ~1.5 ATR; trend exits by time or a trailing stop.

Paper trading. Not financial advice.