Short, dated notes from the desk. Each one comes with the specific observation that would force us to walk it back. Some get shipped into tools and playbooks. Some get retracted — and we keep those visible too.
Every research note starts as a draft and ends in one of three places: shipped into a tool or chapter, kept active with a live falsifier, or retracted in public.
Frame the edge as a specific, dated claim. “X happens Y% of the time when Z.” Vague claims don’t make it onto the notebook.
Define the observation that would force us to walk it back. If we can’t name one, the note is rejected on intake.
Promote a draft to active once the in-sample work is done. Active notes run live for a defined window with the falsifier on the wall.
Notes that survive get shipped into a tool, playbook, or chapter. Notes that fail get marked retracted and stay visible — we don’t quietly delete.
Filter by status or track. Retracted notes stay in the index — we don't quietly delete what didn't hold up.
On the desk, dropping decision-surface latency below ~250ms shifted users from re-pulling charts to acting on first-render data. Above ~400ms, users second-guess the tool. Below ~150ms, returns flatten — diminishing utility.
If session telemetry shows no behavior shift across the 150–400ms band over a 4-week sample, the engineering target gets relaxed.
A simple regime classifier (VIX bucket × value-area context) feeding a multiplier into fixed-fractional sizing beat the unmodified 1% rule on Sharpe across an 18-month walk-forward, primarily by under-sizing during chop.
If forward Sharpe improvement compresses to < 0.10 over 60 sessions, the multiplier gets shipped as opt-in only, not default.
Adding a 1-sigma VWAP band as a hard filter on counter-trend entries cut sample size by ~40% and lifted average expectancy meaningfully across two ticker baskets. Tradeoff: fewer trades, fatter average outcome.
If forward sample expectancy regresses to within 0.05R of the unfiltered version over 200 trades, the filter is dropped.
Members who journal within 90 minutes of session close show measurable mistake-class drop-off in 3–4 weeks. Members who batch-journal weekly take 8–10 weeks to reach the same delta. Same template, same prompts.
If a 12-week cohort shows no statistically meaningful difference at p < 0.10, the recommendation gets softened to optional cadence.
On 5-minute charts, when the daily trend is up and a TTM Squeeze release occurs, the resolution prints upward in ~7 of 10 cases over a 90-day rolling window. The asymmetry collapses near macro events and inside the 60 minutes pre-FOMC.
If the alignment falls below 60% for 30 consecutive sessions outside event windows, demote to confirmation-only signal.
Replaying 2024–2025 trades with a hard 3% daily loss cap and an unchanged position-sizing model cut weekly P&L variance more than halving size with no cap. Caps win on tail control; sizing wins on average return.
If a regime appears where the cap clips materially profitable revenge-trade recoveries (top decile day > +6R after a -3% morning), revisit.
Opening-range breakouts entered between 09:30 and 09:50 ET hit a 2R target ~38% of the time. Breakouts entered after 09:55 ET hit 2R ~21% of the time across the same sample, even when triggers look mechanically similar.
If the post-09:55 cohort recovers to ≥ 30% 2R hit-rate over 60 sessions, the cutoff is widened or removed.
Across two years of intraday ES data, retests of the developing POC mean-revert ~63% of the time when VIX < 18 and the prior session closed inside its value area, and trend-continue at a similar rate when VIX ≥ 22 and price closes outside value.
If 12 weeks of forward data show < 55% directional bias in either regime bucket, the heuristic gets demoted to discretionary input only.
Post-earnings drift trades on large-cap names underperformed expectations by a wide margin in the 2024–2025 sample. We retracted this as a primary edge in member playbooks; remaining mentions are explicitly historical.
Re-instate as an active edge only if 2 consecutive 6-month windows show ≥ 0.30R median drift on filtered candidates.
A short list of questions the desk hasn't answered. If you have data, opinions, or a way to falsify these faster, we'd like to hear about it.
Current sample is 90 days. We want 240 sessions and explicit pre-FOMC / pre-CPI carve-outs before promoting to a primary signal.
Members who journal within 90 minutes of close also tend to size smaller. We need to disentangle which lever is doing the work.
The lift looks real on two ticker baskets. The honest answer is we haven’t stress-tested it on low-liquidity names where VWAP itself is noisy.
We saw the 250ms inflection on the desk web client. Mobile and desktop NinjaTrader users may have a different sweet spot. Open.
If you've replicated, broken, or extended one of these notes, we want to see it. If you'd rather just keep reading, the long-form blog and the open primitives are next door.
Found a regime where one of these notes breaks? Send the data — we’ll review it and credit you in the revision.
The R-multiple math, regime classifier, and journal schema are public. Fork them, plug them into your own pipeline.
Notes here are short by design. The blog has the full essays, teardowns, and chapter previews.