Spent the whole day on my paper-trading bot, and it turned into an accidental case study in letting backtests veto your own ideas. Five features built today; two got deployed, three got shelved — each decision made by a number, not a feeling.

The day started with performance. The walk-forward backtest was recomputing indicators on a growing slice every bar — classic accidental O(n²). Two fixes: compute vol on a fixed 22-row tail (verified value-identical to 1e-15, just pandas rolling float noise), and read precomputed indicator rows instead of re-running the whole pipeline per bar. Full 11-year, 18-symbol backtest went from minutes to seconds. That speed mattered later: once a backtest costs 40 seconds, you stop arguing about ideas and just measure them.

Then the fun part. My options-wheel strategy loses to buy-and-hold exactly on trending stocks — covered calls cap every rally. So: a “router” that predicts P(rally > 8% in 30 days) per symbol and holds shares uncapped instead of wheeling when it’s confident. First predictor was LightGBM on technical features. Backtest verdict: worthless — best case +$1.4K on a $46K deficit, pure noise, holdout AUC ~0.5. Second predictor: Google’s TimesFM, zero-shot forecasting the 30-day SMA direction. Verdict: recovered $25–29K of the deficit across a stable threshold plateau, and it held up in a 2024+ out-of-sample window (with the honest caveat that a foundation model’s pretraining may brush your backtest years — the only clean data is live paper trading going forward). Same routing rule, different predictor, opposite outcome. The mechanism was never the problem; the prediction was.

The deployment pattern I landed on: mechanical proposes, LLM disposes. The backtested rule generates the signal; the LLM overseer approves by default and vetoes only on context the models can’t see (earnings, halts). Every veto is logged so I can measure whether the LLM’s judgment helps or hurts versus the pure rule. No prompt-engineering hope allowed to silently replace a measured policy.

The afternoon delivered the best negative results. I built a portfolio-level backtest — all symbols sharing one cash pool — to test capital-allocation policies. My “obvious” score (premium per day per collateral dollar) lost $47K versus dumb first-come-first-served at small capital: fat premium density is a high-IV signature, so the score systematically funded the riskiest names and starved the best one. A v2 with a per-symbol edge prior recovered half the gap in the full window, then a 2023+ window flipped the entire ranking. Conclusion by minimax regret: neutral ordering, plus one strictly-dominant tier (cash-freeing signals always jump the queue). Two days of “improvements” correctly reverted by their own measurements.

One more role made the cut: buy quality dividend ETFs at the panic close (VIX ≥ 30), sell when recovery confirms. 19 real episodes since 2015: 15–16 wins per ETF. I doubted my own entry rule — buying into a crash feels like knife-catching — so I tested waiting for the first calm bar instead. Waiting cost 40% of the profit. Panic prices are the discount; by the calm bar, the bounce already happened. Also tested and excluded: long treasuries (6/19 wins — flight-to-safety is expensive at panic) and gold (profitable but only because of the recent gold bull).

Ops bugs of the day, both found by just starting the container and watching: a market-closed day made restart: unless-stopped hot-loop the process every 5 seconds, spamming a Slack notification per cycle (fix: sleep until the next 9am ET check instead of exiting). And a YAML folded scalar (command: >) had preserved a newline before a more-indented continuation line, silently splitting my shell command in two — the tee that was supposed to persist logs had been connected to an empty statement for weeks. Compose commands are lists now.

The weekend ends with an A/B experiment armed: two scanner instances in tmux, main branch versus feature branch, identical starting ledgers, same LLM, isolated state. Monday’s open decides the next argument with data.

Lesson of the day, twice confirmed: the ideas that survived measurement were not the ones I would have bet on in the morning.