Breakthrough Preferences oil

Sell model: 100% trigger accuracy, zero false sells, CVaR99 -24.6%

breakthroughoilquant-finance
Key Metric
Before
No systematic sell discipline : manual exits, no trigger...
After
11-trigger sell model, 100% accuracy, 0 missed sells, 0 f...

Zero false triggers. Zero missed sells. A Brier score of 0.0024 : near-perfect probabilistic calibration. The sell model is the risk management layer the oil model was missing since day one.

Context

A pricing model that can tell you when to buy but not when to sell is only half a trading system. Before v16, position exits in the oil model relied on manual judgment: ad-hoc decisions based on current price relative to entry, gut feel about geopolitical developments, and informal profit targets. This meant that even accurate buy signals could be degraded by poor exit timing. The model’s edge on entry could be given back on exit.

The experiment designed a systematic sell trigger framework from scratch, motivated by the question: what are all the conditions under which a rational trader should exit a crude oil position?

What Changed

11 specialized sell triggers were defined and implemented, organized by the type of exit condition they address:

Take-profit triggers fire when the position has achieved its probabilistic target. P&L stop triggers limit downside on positions that move against the thesis. Conviction hold logic maintains positions through normal volatility when the fundamental thesis is intact. Hopeless position detection exits positions where the entry thesis has been falsified by new information. Edge erosion exits when the model’s confidence in the position’s expected value drops below a maintenance threshold. Dust triggers exit positions too small to justify continued management overhead. Stale lottery exits options-equivalent positions that have lost time value without moving. Portfolio circuit breaker limits aggregate exposure across concurrent positions.

CVaR99 measures extreme tail risk : the expected loss in the worst 1% of scenarios. Reducing it by 24.6% means the worst-case outcomes are significantly less severe.

Impact

Before: no systematic sell framework. Manual exits. No quantitative trigger discipline. After: 11 triggers covering every relevant exit condition. 100% trigger accuracy across all 3 quant audit rounds. Zero false sells (no premature exits). Zero missed sells (no positions held past their trigger). Sell Brier score of 0.0024. CVaR99 reduction of 24.6%.

A Brier score of 0.0024 is near-perfect probabilistic calibration. The sell model produces sharper exit probability estimates than nearly any forecasting system at this scale. Combined with the CVaR99 reduction, this transforms the risk profile of the entire trading system.

Source