Position Sizing 101: Why Our Paper Bot Risks 1% Per Trade

Entries get all the glory. Sizing decides whether you survive to trade another day. Our paper bot's worst stop-out so far was −72.4% on a single new-token position — and the run is still alive, because that position was a fixed $25. This is the exact sizing rulebook it runs on. Educational only — how a paper system is built, not a recommendation for anyone's money.

The math: volatility-adjusted, not vibes

For the majors, position quantity is equity × risk% ÷ (2 × ATR), where ATR is the average true range on the 1-hour chart — a standard volatility measure. A choppier coin gets a smaller position; a calm one gets a larger one. Same risk budget either way. Then three hard caps apply: no position exceeds 25% of equity in notional, total deployment never passes 60% (40% cash reserve), and no more than 7 positions at once.

New tokens get different treatment: a flat $25 per position because their volatility is extreme and unknowable — ATR math breaks down when a token can lose 90% in an hour. The bot's own data proved the point: the universal 1-hour exit test showed non-band holds cratering to a −98% median. Small fixed size is the only sane answer.

The kill-switch ladder

• −3% in a day → no new entries for 24 hours
• −5% in a day → full halt for 24 hours
• −10% in a week → pause for 72 hours
• −15% total → FULL STOP, manual reset required
• >5 consecutive losses → halve position size until a win

Why 1% and not 5%?

Because fees and streaks are structural, not bad luck. On Kraken, taker fees are 0.80% per side — 1.6% round trip — so sub-1-hour scalping is mathematically dead before it starts. And every system hits losing streaks; at 1% risk, ten straight losses cost ~9.6% of equity. At 5% risk, ten straight losses cost ~40% — and you need a 67% gain just to get back to flat. Sizing is what turns a strategy's edge into something that can survive variance.

Keep/kill, no tinkering

The bot runs a 30-day evaluation with pre-written verdicts: KEEP if profit factor ≥ 1.3, expectancy > 3.2% per trade, max drawdown < 10%, and ≥ 20 trades; KILL if profit factor < 1.0 after 30+ trades. The rule that matters most: no re-optimizing on the same data. Tuning until the backtest looks good is how you manufacture a beautiful lie.