Trading: a strategy desk's decision journal

Why didn't the desk fully exit NVDA on the March 2025 death cross, and what brought the position back to 100%?

On March 27, 2025 the 50-day average crossed below the 200-day at $111.43. The desk cut to 25% of target size instead of exiting. Single-name death crosses are noisier and win less often than index-level ones, so the cross became a size-down trigger paired with an ATR-based stop. On August 13, a short-horizon golden cross fired while the primary system was already back in risk-on. Corroborated, the desk returned to 100%.

The answer joins two entries five months apart. The journal is a running narrative: two signal events, one backtest, three sizing decisions. State changes over time: the case for continual reasoning under uncertainty.

Where flat summaries slip

What would cut the position back down? A rule stated once in August: hold size until a new 50/200 death cross, or a close more than 1x ATR below the 200-day average, whichever comes first. It is load-bearing for every question about what happens next. It survives as its own belief.

Was the August golden cross the signal the backtest measured? No. The backtest measured the 50/200-day system. August's signal was a separate 20/50-day variant. "Golden cross" names both: an ambiguity, easy to conflate. Kept distinct.

Does the desk think the strategy has an edge? Both sides survive. The backtest verdict: better risk-adjusted return, Sharpe 1.05 against 0.89 for buy-and-hold. The closing entry's open risk: position sizing may carry more of the return than the signal itself. An unresolved risk stays a belief, not a footnote.

What changed for trading

Nothing domain-specific. Two real trading documents, a trend-following journal and a note on ATR position sizing, ran through the generic five lenses as test material. They exposed four general bugs:

• The same fact, worded two ways by two lenses, counted twice.
• The explanation lens invented reasons the document stated elsewhere.
• Dense blocks of numbers were only partly read.
• Headings and timestamps were mined as if they were facts.

Each fix went into the shared pipeline once. Every domain inherited it.