Brian Shannon's intraday VWAP bounce: the full test report

What we granted, and what we actually tested

Granted, not measured: Brian Shannon teaches a specific touch, reclaim, new-high entry and higher-low stop. The visual check supports detecting that pattern. We grant the rule its stated rationale; the discretionary choice of stocks and tape reading are outside this test.

Tested: One intraday session-VWAP bounce, with one-minute structural stops, on 28 fixed liquid survivor stocks from 2022-04-26 through 2026-06-11. 8,636 simulated trades; fixed-fractional 1% risk; no overnight positions. Session VWAP resets every morning. This is distinct from multi-day event-anchored VWAP.

§1 · The locked pre-registration

Committed 471ae34 on 2026-06-11, before the verdict run — the wager you cannot move after seeing the cards.

The falsifiable question

Can this encoded intraday bounce clear an absolute after-cost cash hurdle AND outperform matched random entry in the same eligible stocks?

• A yes looks like: Net annualized return at least cash + 1.0 percentage point, positive aggregate net P&L, positive net expectancy, and return above the 95th random-entry percentile.
• A no looks like: Failure of either gate is a fail. Selection skill with sub-cash returns is recorded distinctly; it cannot be promoted into an absolute pass.
• Ambiguous residue: Near a threshold, investigate a named plausible false-negative artifact before a terminal call. This absolute miss is not near the registered hurdle.

Our stated prior: Friction likely decisive

The registered floor recorded that the visual check "confirms detection FIDELITY, not edge" and that "the modest setup quality makes FRICTION the likely decisive factor". This was a qualitative prior, not a registered numerical probability.

Primary configuration: pre-run and run-stage decisions

Untested scope limits — when they were identified

1. Multi-day event anchors were identified as a reopening test after the result, not named in advance. Discretionary stock selection and tape reading remain outside this mechanical test.
2. Gap-up handling and confluence with other indicators were identified as explicit scope limits in the later source review, after the result; they were not named in advance. Choppy-market sizing adjustments were not tested.
3. A historical constituent universe with failures and delistings; deep bears outside this data window.
4. Slower-chart stop reconstruction was identified after the result in the September source review, not pre-registered; it is the strongest unresolved objection.

The goalposts — all required to proceed

1 of 2 gates passed; the pre-registered conjunction fails. Gate 2 uses the priced runs, not a free-fill comparison.

§2 · The headline arithmetic

§3 · Was it the path? The stress tests

The entire cost grid follows. R means the entry risk unit, the distance from entry to the average of the two initial stops. A trade can cost a large fraction of R even when its cost is small relative to the stock price.

The cash-relative -104.49 figure is a difference in annualized percentage points. It is not a literal account loss greater than 100%. The model account shrinks toward zero under 1% fractional risk; cash compounds at 4.49% annually. The first threshold crossings below were recorded in a dated post-result rebuild:

All 28 of 28 stocks have negative total net R at the registered price. NKE has only 7 trades. The registered ex-top test drops NKE, the least negative stock, and remains negative. The separate September 19 post-hoc test drops the worst stock, LLY: 8,092 remaining trades average -0.5237 R. That result was selected after seeing the outcome; its random-entry gate was not rerun.

All 5 of 5 calendar regimes have negative net mean R:

§4 · The controls

Fidelity before profit: the first detector produced 13 real bounces in a 36-chart sample. Tightening the touch tolerance and excluding the unstable opening bars reduced signals from 14,799 to 8,722. The second sample contained 36 of 36 real bounces, but 31 were marginal and five were clean. We graded them ourselves, with future prices hidden. Acceptance of that disclosure was not independent re-grading. The earlier 33-marginal transcription was corrected to 31 before the result.

Signals to trades: 8,722 signals became 8,636 trades. All 86 of 86 exclusions were signals arriving while a position in that stock was already open. This bridge was counted in the September 19 rebuild.

Random-entry controls: 2,000 portfolios, seed 20260611, from pools of 1,500 structural twins per stock. Per-name trade count, long direction, eligible sessions and the bracket machinery are shared. Holding periods emerge from that machinery; their distribution was not forced to equal the system's. All controls pay the same registered fill cost. Ranking above these controls is evidence of selection value inside this comparison, not proof that the account can pay its bill.

Cost identity: over 8,636 trades, mean(0.10% of entry / risk unit) = 0.545929 R, versus the grid drop 0.545937 R. The median-trade cost is 0.405 R; the mean is higher because of the tight-stop tail. Splitting an exit into thirds does not increase total notional: the earlier explanation attributing this gap to exit granularity was wrong.

Intrabar ambiguity: 23.77% of trades contain a same-bar stop/target ambiguity. The lab's optimistic one-R-per-affected-trade bound is -0.291 R per trade, still negative. This arithmetic bound is not a full alternative fill simulation.

Sanity alarms: 0 recorded alarms; the run trades in the bear, does not depend on one trade, is not mostly closing at the bell and produces thousands of trades. These checks answer those specific questions; they do not prove an error-free simulator.

§5 · The strongest objection and what remains untested

The strongest unresolved objection is the stop timescale. Shannon describes moving from one-minute to two-minute and then five-minute charts. We kept one-minute pivot lows throughout. A wider stop changes the cost measured in R, the central mechanism here. The slower-chart rerun has NOT RUN; it is the first reopening criterion. This objection was sharpened after the original result, so it is not presented as a completed pre-registered stress test.

Two favorable-data limitations: the fixed universe contains survivors, and the stage-2 uptrend filter uses the same day's closing information, unavailable earlier in that day. The random-entry controls share the filter. This timing look-ahead limits tradability and may alter selections; a causal-filter rebuild has not been run. Neither a shared bias nor a failed net result makes the setup operationally validated.

§6 · Their own terms

Source: TraderLion, The AVWAP Trading Indicator | Secrets and Setups | Brian Shannon, CMT. The entry-rule passage is at 48:01. The taught sequence is touch, reclaim, first new high after the touch, and stop under the relevant higher low. Our numeric choices and simulator conventions are distinct from that teaching.

The following 26 choices are the rule-to-code ledger. Detector and floor choices were frozen at the June 11 ratification. Simulation conventions 18 through 26 were appended during the result-stage implementation under the registered allowance; they are not represented as earlier separately committed decisions.

§7 · The complete ledger — all 8636 rows