A forensic audit of the owner's own trading tape
- Window
- 2024-08-23 → 2026-08-20
- Method
- three independent analysts plus a benchmark pass
- Tape ends
- 2026-06-10
- Active days
- 588
£132,124 realised on £171,870 of capital while it was actually at work — +76.9%, 42.5% a year over 588 active days.
Spread across the full 728-day calendar the same P&L is +93.7% on £141,059, 39.3% a year.
Take away the top names one at a time
LEAVE-ONE-OUTStated in full at § 1 · The record in six lines, row 3 — The record is its right tail.
stock + options + excluded
1,288 + Revolut/T212 equity 827
3,596 crypto, 1,163 CFD, 104 forex
£0.0000
1,932 of 2,115 fills (91%)
The record in six lines
The headline is the beginning of the audit, not its conclusion. Each line carries its confidence mark.
The headline
£132,124 realised. Measured over the days his capital was actually at work — 588 of the window's 728, in two spells — that is +76.9% on £171,870 of deployed capital, 42.5% a year. On the full calendar the same £132,124 is +93.7% on £141,059, 39.3% a year — same P&L, same capital at work, just a clock run 140 days longer: 69 of them completely flat, 71 of them a stale lot book with no fill behind it. Across three spell definitions and three gap thresholds the annualised rate lands at 40–48%; across four capital denominators the cumulative band is 80–98%. Read as per year, he made the middle of his own "40–50%" claim, not the bottom of it. Method in §1b.
+76.9%Versus what
The like-for-like test is to hold the index only on his own spells, and it moves both sides. On the calendar SPY returned +38.9% and QQQ +49.7% (total return, dividends reinvested) — 22.4% a year for QQQ. Over his 588 active days QQQ returned +46.2%, or 26.6% a year: the market's good stretches largely were his active stretches, so restricting to his spells flatters the benchmark more than it flatters him. His annualised edge over QQQ is therefore stable at +15 to +22pp on every basis (+16.9pp on the calendar), worth £53k–£62k in cash. But the max drawdown on his own cumulative realised curve was −£62,927 (2026-02-20), a 47% giveback of peak gains, against QQQ's −22.9%. Roughly QQQ-sized pain for roughly 2× QQQ's return.
+46.2%The record is its right tail
7 of 128 positions produce 80% of the P&L. Leave-one-out: remove ACHR → 71.9%; remove ACHR + CRCL → 55.5%; remove ACHR + CRCL + SHOP → 44.4%, which is 5.3pp below QQQ. Cap every winner at his own median winner (£1,220), keep every loss at full size, and the two years become −£12,480.
7 / 128The entry edge is real, and it is thin
Buys land −16.96% below the stock's own trailing 60-day high; sells at −13.46% — a +3.51pp round-trip edge, on £6.4M of buys, levered by ~24× annual turnover. Not a big idea. A small, repeated one.
+3.51ppThe named theses are not the earner
In-window: GLP-1 −£1,379, memory +£264, Google-AI +£13,856 — under 10% of the total, on ~5% of capital. The actual engine was eVTOL (ACHR + JOBY, £36,830) and crypto-equity proxies (CRCL/COIN/HOOD/CLSK, £29,724): high-beta, high-retail-attention US names.
£36,830The edge decayed as it scaled
Year 1 +78.9% on £87k average capital; year 2 +32.5% on £195k. Realised £ per £100 of stock bought: 2024 £26.87 → 2025 £4.06 → 2026 £0.46. The tape stops dead at 2026-06-10.
£0.46Returns by active trading spell
The same realised P&L, measured only while capital was at work and benchmarked on those identical days.
Definition. A day is active when the FIFO lot book has open inventory (open cost > £1) or a fill occurs. A spell ends after 14 consecutive inactive calendar days. Return per spell = realised P&L ÷ time-weighted deployed capital inside the spell; annualised as (1+r)^(365/active days). Benchmarks are the same daily total-return series used everywhere else, sliced to each spell's exact dates.
The spells. The window contains exactly two, separated by one real break — 69 completely flat days, 2024-12-19 → 2025-02-25, with zero cost basis and zero fills.
Two spells, one pooled rate
TAPE-TRUNCATED EXPOSURE| Spell | Start | End | Active days | Capital | P&L | Return | Annualised |
|---|---|---|---|---|---|---|---|
| Spell 1 | 2024-08-23 | 2024-12-18 | 118 | £32,150 | £20,765 | +64.6% | 367.0% |
| Spell 2 | 2025-02-26 | 2026-06-10 | 470 | £206,949 | £111,359 | +53.8% | 39.7% |
| pooled | 588 | £171,870 | £132,124 | +76.9% | 42.5% |
Annualised on active days: his 42.5%, SPY 17.4%, QQQ 26.6% — an edge of +15.9pp a year. Return per active trading day: 13 bps. Per-spell annualised figures are deliberately omitted; annualising a 118-day spell gives 367%, which is arithmetic, not a rate.
The tape vs QQQ vs SPY on the same active days
IDENTICAL DAYSLenses A and A′ are completely insensitive to the threshold — there is only one inactive stretch in the window and it is 69 days long, so 7, 14 and 30 all cut in the same place. Only lens B responds, and it moves the annualised figure by 4pp across the whole 7–30 range.
Return bought with drawdown
The return clears the benchmarks. The path to it was far less forgiving, and the profit sits in a very small right tail.
Return bought with drawdown
SAME RECORD · DIFFERENT PAINStated in full at § 1 · The record in six lines, row 2 — Versus what.
The sources state no drawdown figure for SPY, so that side of the axis is left blank rather than plotted as zero.
Where the profit actually sits
128 POSITIONSStated in full at § 1 · The record in six lines, row 3 — The record is its right tail.
P&L by theme
NAMED VS ACTUALThe strategies he actually ran
Named from the tape, not from self-description. Each mechanism carries its evidence line.
Dip-tranche accumulation
Many small, near-identical clips into falling prices. Evidence: 553 follow-on buy tranches across 48 names — 58.8% priced below the previous tranche, 60.8% below running average cost. Against market data, the median stock is −2.37% over the 5 sessions before he buys (baseline for the same names over the same windows: +0.39%) and −3.64% over 20 sessions (baseline +1.89%), sitting at the 33rd percentile of its trailing 20-day range.
He is early, by 2–4 points
those same buys go on to return a median −3.11% over the next 20 days against a +1.32% baseline. He provides liquidity to a 3–5 day decline and gets paid by the bounce — unless the decline keeps going.
Sell into strength, in scaled tranches
The better half of the system. Evidence: sells strike after a median +4.15% 5-day run, at the 69th percentile of the 20-day range, and are followed by −2.42% over the next 20 days (baseline +1.32%). Exits are batched more than entries — 68 episodes exiting in 4+ fills carry £103,438, 70% of the whole book's P&L.
No disposition effect
Genuinely unusual, genuinely positive. Evidence: cost-weighted holds — winners 6.3 days (n=92) vs losers 7.3 days (n=49). He does not cut winners and marry losers.
Same-name rental
He does not hold names; he rents them, repeatedly. Evidence: 59% of consecutive same-symbol episode pairs are re-entered within 30 days; symbol-days containing both a buy and a sell went 1/89 (2024) → 46/497 (2025) → 123/530 (2026). Median hold by exit year: 39.1d → 10.0d → 3.5d. The regime break is the broker, not the market: Revolut era 50 closed episodes, 24.3-day median hold, 72% win, +£91,513; IBKR (from 2025-07-14) 90 episodes, 2.7-day median hold, 62% win, +£37,452.
Flat clips, no pyramiding
median tranche-size ratio 1.03× when adding higher, 0.98× when adding lower. Position size is a function of tranche count, not conviction. Book widened anyway: largest position fell from 59% of book (2023) to 27% (2026), average positions held 3.9 → 15.0.
Ladder in correctly, capitulate out
The failure mode, and it is the same shape every time: a residual stub left alone while it falls 25–60%, then dumped within a few percent of the low. Evidence: KLAR 302.50 sh held 129 days — 90 of them with no KLAR fill at all (2025-11-21 → 2026-02-19) — sold 2026-03-27 at $12.52 against a $12.27 low, then +66%; NVO 250 sh held 72 days, 71 of them untouched (2026-02-03 → 2026-04-15), sold at $39.80 against a $38.52 low; OPRA liquidated over 8 fills, 2,840.6 sh, 2026-01-15 → 2026-02-02, at $13.42–$14.60, now $18.96.
Six positions that teach the whole record
Two wins that reveal the engine, one patient campaign, one natural experiment, and two broker-specific failure modes.
ACHR
+£30,724Across both brokers = 23.3% of the two-year total
All of it in-window (the first fill is 2024-11-21, after the window opens). The Revolut leg is the earner: 131 fills, 2024-11-21 → 2026-02-10, +£31,641. The IBKR leg gave £917 of it back over 44 fills, 2026-02-02 → 2026-05-06. Buys at −23.7% below the 60d high, sells at −9.9% — a 13.8pp spread, the widest in the book. October 2025 alone: +£17,651 on 13 fills, selling 8,333.9 sh at a quantity-weighted $11.76 into a two-year high of $13.03. It is $6.30 today.
SHOP
£15,733In-window, and the record's one genuinely patient trade sitting inside a lot of trading
Two campaigns, two brokers, 54 fills. He first bought 2022-10-10 (259 sh @ $27.00) and sold the lot a month later. The core position is two January-2023 lots — 94.4 sh @ $36.79 and 237.9 sh @ $42.04, £10,414 of cost — bought after SHOP fell from $169 to $25.67, and those two lots have a cost-weighted hold of 465 days, about 15 months. That is the patient trade, and the only stretch he truly left alone is the 340 days between 2023-01-23 and 2023-12-29. From there he traded it: a first trim 2023-12-29, 200 sh out on 2024-03-25, then 7 buy tranches and 10 sells between 2024-04-09 and 2024-12-04, with individual lots held anywhere from 2 to 176 days. The often-quoted 693 days is the span from that first January-2023 buy to the last 2024 exit — it contains 22 fills, and is not idle time. Then in 2026 he rented the same ticker on a stopwatch: 30 fills, 2026-02-13 to 2026-05-14, holds of 1 to 29 days, £123,335 cycled for +£6,042. All-time the Revolut campaign made +£17,851, +98.8% measured against £18,067 — the highest FIFO open cost the Revolut SHOP book ever carried, on 2024-08-06 — though that ratio blends the 2022 round trip into the 2023–24 campaign's peak capital; the 2023–24 campaign alone is +91.9%. Cost-weighted hold, defined as Σ(FIFO lot cost × days held) ÷ Σ(lot cost): 203.8 days over the Revolut lots all-time, 49.9 days over every SHOP lot ever, 18.3 days over the lots closed inside this window.
GOOGL
+£12,753The thesis that worked and the exit that didn't
Built through the March–May 2025 drawdown across 23 buys at a quantity-weighted $158.92, including 339.1 sh on 2025-05-07 at an average $151.76. Exited over 27 sells at an average $180.59, the last on 2025-08-19 at $201.47. GOOGL closed $340.67 on 2026-08-20 — held instead, the 975.3 sh would have been worth $332,244 against $176,126 of actual proceeds, so the exit forfeited $156,118, more than his entire two-year realised stock P&L.
RDDT
+£11,091 / −£6,550The cleanest natural experiment in the data
Same name, same dip-buy entry quality. Revolut: 25 fills, 31.8-day cost-weighted hold, +£11,091 (+36.9% on £30,088 of peak open cost) — £10,175 of it in-window, the rest earned before the window opens. IBKR: 78 fills, 2.5× the peak capital (£74,205), 7.3-day hold, −£6,550 (−8.8%). The buy/sell spread vs the 60d high collapses from +16.5pp to +2.1pp. The edge is in the holding period, not the stock selection.
COIN
−£12,876The churn machine
258 fills in 190 days on one ticker; 118 buys, 84 of them below running average cost; £1.36M of buy notional = 20.2% of all buy notional ever transacted (£1.49M and 22.0% counting the Revolut leg too). Peak commitment somewhere between £134k and £166k depending on whose lot book you use — i.e. roughly one entire book in one ticker. February 2026 alone: −£30,209. Edge 2.1pp; too thin to survive that size. The name was not the problem: the same ticker made +£11,998 at Revolut, so COIN across both brokers nets to −£878.
CRCL
−£6,720The purest capitulation
17 consecutive strictly-lower buys, $198.95 (2025-07-22) → $83.33 (2025-11-13), −58.1%, $93,377 deployed for 856.8 shares — eleven of the seventeen crammed into five November trading days, five of them on 2025-11-12 alone. He then sold the whole accumulation starting 2025-11-21 at $68.40, 24.9% below his own $91.09 average cost. CRCL is $87.98 today. He was right about the level and wrong about his own tolerance.
The hole and the climb out (CRCL / COIN)
Both names, both brokers, and the dated mark-to-market path: a real recovery that splits between market rebound and trading contribution.
From peak, through the hole, to the exit
RUNNING P&L · BOTH NAMESThe two paragraphs above tell half of each story. Each name was traded at both brokers, and the halves point in opposite directions. CRCL: Revolut −£6,720 (62 fills) against IBKR +£29,778 (206 fills) = +£23,058. COIN: Revolut +£11,998 (53 fills) against IBKR −£12,876 (258 fills) = −£878. Together, 579 fills for +£22,180. Neither name was a net loser, and the 206-fill CRCL leg at IBKR is the largest single position the record had not reported.
The hole is real and it is dateable. Marking both books to daily closes, the combined CRCL+COIN position bottomed at −£41,444 on 2026-02-12 — realised −£15,665 plus unrealised −£25,778, against £118,109 of cost then deployed, a 35% hole. The largest peak-to-trough fall was £54,211 (+£12,767 on 2025-10-27 → −£41,444). By 2026-05-28 the pair stood at +£22,169: a £63,613 swing in 106 days.
What he did. Digging in (2025-10-27 → 2026-02-12): 175 fills, £145,721 of net new money added as the position fell, in near-identical ~£10k clips on essentially every trading day, no cut. Climbing out (2026-02-12 → 2026-05-28): 311 fills, the cash flow reversed to +£180,497 out, 171 sells against 140 buys, 29 same-day round trips — and +£29,650 of the recovery landed in May 2026 alone, on £959,719 of buys.
How much of it was him. Freeze the book at the bottom and mark it forward: doing nothing would have ended at −£6,380, so the market closed 84.6% of the hole by itself. His trading added +£30,666 on top (price-based; +£28,549 via the file's realised P&L), which is what turned break-even into +£22,169. The two halves sum exactly — £35,064 market + £28,549 trading = the whole £63,613 climb, i.e. 55% market, 45% trading. The context is that this was sector-wide — over the dig CRCL −60.5%, COIN −61.0%, HOOD −51.2%, CLSK −53.9% against QQQ −4.4%; over the climb CRCL +91.1%, CLSK +94.8%, COIN +29.2%, QQQ +22.5%.
Is there a strategy in it? One candidate, not promotable: sector-capitulation scaled exit. The climb-phase sells do carry a real forward edge — they strike at the 81st percentile of the trailing 20-day range and are followed by −7.23% over 10 sessions against a −0.17% baseline (n=171). But the same phase's buys are followed by −6.81% against that baseline (n=140), almost exactly cancelling it; across the two names the price-based legs point in opposite directions; and leave-one-out kills it — remove CRCL and the pair is −£878. It is one name, at one broker, in one V-shaped recovery, with no stop and no position cap anywhere in the tape, at a sizing that put roughly a third of the book in one ticker. The single most valuable thing he did was not sell at the bottom, and that is a temperament, not a mechanism. Full episode, tape, event study and the mechanism brief in the plain-English format: owners-record-crcl-coin.md.
Each broker tells the other half
TWO-NAME RECONCILIATION| Name | Broker | Fills | First | Last | Realised |
|---|---|---|---|---|---|
| CRCL | Revolut | 62 | 2025-06-06 | 2025-12-09 | −£6,720 |
| CRCL | IBKR | 206 | 2025-11-19 | 2026-05-14 | +£29,778 |
| CRCL | both | 268 | +£23,058 | ||
| COIN | Revolut | 53 | 2024-10-23 | 2025-10-06 | +£11,998 |
| COIN | IBKR | 258 | 2025-11-19 | 2026-05-28 | −£12,876 |
| COIN | both | 311 | −£878 | ||
| both | both | 579 | 2024-10-23 | 2026-05-28 | +£22,180 |
How it would trade
- Trigger. A high-beta name down ≥50% from its 60-day high, where the fall is sector-wide, not idiosyncratic — require ≥3 named peers also down ≥40% over the same window while the index is down <10%. From the tape: CRCL −60.5% and COIN −61.0% with HOOD −51.2% and CLSK −53.9%, against QQQ −4.4%.
- Add ladder. Flat clips, no pyramiding: median £10,107 per clip at a CV of 0.52, added on roughly every down day (median 1 trading day between fills). The tape has no cap — the ladder ran to 17 consecutive lower buys at Revolut and 87 buys through the dig. A tested version must impose one; the tape cannot tell us what it should be.
- Sizing. His peak single-name commitment was £134,424 (COIN, 2026-01-29) against a book that peaked at £404,842 — roughly one third of the book in one ticker. That is the sizing the result was earned at, and it is also the reason the strategy is dangerous.
- Exit. Scaled sells into strength: 171 sells against 140 buys in the climb, striking at the 81st percentile of the trailing 20-day range, followed by −7.23% over 10 sessions against a −0.17% baseline.
- Stop. None exists in the tape. This is the missing rule, and the honest reason the mechanism is not testable as run.
When it wins · when it loses
- WINS — V-shaped, sector-wide drawdowns that mean-revert hard inside ~3 months. In this episode the complex bottomed together and CRCL rose 91.1% and CLSK 94.8% off the low.
- LOSES — an idiosyncratic fall (the company really is broken; no peer confirmation), or an L-shaped sector decline that never rebounds. It also loses to its own add-ladder: the buy leg cost £74,384 of price-relative value and climb-phase buys were followed by −6.81% at 10 sessions against a −0.17% baseline.
- RUIN — at the bottom he was £118,109 deployed and £41,444 under water with no stop and no position cap. A further 30% leg down is not survivable at that sizing. The strategy's headline result is conditional on having had the capital and the nerve to not be forced out, and that condition is unmodelled.
What we're looking for
- On a corpus of names that fell ≥50% with ≥3 confirming sector peers, does a scaled-exit rule beat buy-and-hold from the trough, after costs? Prior from the tape: +£30,666 — on n=1 name, which is not evidence.
- Does the sell-timing edge survive conditioning on the 20-day range percentile? His sells sat at the 81st percentile against a 57th-percentile baseline. If high-range fills in a mean-reverting tape are generically followed by falls, the edge is an artefact.
- Should the mechanism be exit-only? The tape says yes: the add-ladder destroyed value on a forward basis. Test the exit rule against a passive entry before testing any ladder.
- Name the payer. The lab's own first gate: who is on the other side and why do they keep paying? The plausible answer is forced sellers and margin liquidations in a levered, retail-heavy sector during a crash. It is not evidenced anywhere in this data. By the lab's rule — no named payer, no test — this does not yet qualify.
- Kill criterion. If the exit rule does not beat hold-from-trough on ≥60% of a ≥30-name corpus with a 25–50bps cost model, it dies at the free look.
What the lab does with this
Five things carry forward. Each inherits a mechanism from the tape and a constraint from the audit.
Three ideas were promoted to named theses — and the scoreboard is published with them
GLP-1, Google-AI and memory are tracked because he holds the view, not because the tape rewards it. Evidence: as trades, all three together made £12,740 in-window, under 10% of P&L. As assets over the same window: MU +847%, GOOGL +106%, LLY +31%, NVO −66%. Timing dominated selection — MU only went vertical from ~Sep-2025 and GOOGL from ~Jul-2025. Ambition: that the theses eventually pay. Evidence today: they don't; the eVTOL and crypto-proxy trades do.
The thesis-drawdown-entry screen inherits mechanism #1, with the parameters the tape actually implies
Entry depth is the live variable: META's −19.5%-below-high leg made +£10,061; its −9.7% leg lost £3,688; NVO's non-dip entry at −7.9% produced the worst inverted edge in the book (−10.4pp). The screen therefore tests a depth floor vs the trailing 60-day high, a tranche cap (heavy averaging-down — 3–5 down-adds — wins only 40% of the time, n=10), and a mandatory exit rule for residual stubs, which is where every large loss came from.
Hold length becomes a tested parameter, not a preference
The RDDT pair and the broker split give the prior: ~30-day holds worked, ~7-day holds on the same names did not.
The benchmark layer keeps it honest going forward
Every strategy is scored against SPY +38.9% / QQQ +49.7% over its own window, not against zero; every result carries a hold-everything counterfactual per name; every result carries a leave-one-out-3 line, because this record passes LOO-2 and fails LOO-3; and every backtest carries a cost model of 25–50bps on any venue reporting zero commission, because that is the single largest unmeasured number here.
Churn tax is a live gauge
Realised £ per £100 bought (£26.87 → £4.06 → £0.46) is the cheapest early-warning metric in the whole audit and costs nothing to compute daily.
What this audit cannot tell you
Nine limits, in source order and ranked by how much they can move the headline. None is abbreviated.
The capital base is an approximation
Deployed capital = end-of-day FIFO cost basis in GBP, averaged over 728 calendar days. It is deployed capital, not account capital: idle broker cash is invisible, and no market marks exist anywhere in the data. Four denominators give 32.6% to 97.6% cumulative (peak-capital basis to broker-scaled basis); the primary FIFO figure is 93.7%. Against the one broker snapshot available, the tape overstates aggregate open cost by 4.2% — good to ~5% in aggregate, but wrong per symbol (SHOP 250 sh tape vs 80 snapshot; NFLX 35.5 vs 135.5).
"Active" is a definition, not a fact
A day counts as active when the FIFO lot book holds open inventory or a fill occurs; a spell ends after 14 consecutive inactive days. On that definition the window contains two spells and only one real break — 69 flat days, 2024-12-19 → 2025-02-25. Two judgement calls follow. First, the last 71 days of the window carry £22,980 of cost basis with no fill behind them; §1b truncates the series at the last fill (2026-06-10) rather than treating a stale lot book as live exposure — keeping it gives 659 active days and 40.5% a year instead of 42.5%. Second, counting only days he actually traded (a fill within 14 days) gives 532 days and 46.0%. The 7/14/30-day thresholds change nothing for the exposure definition and 4pp for the trading definition. A money-weighted XIRR was computed and is not published: it lands anywhere between 66% and 97% depending on how you treat the £58,991 of in-window sell proceeds that have no recorded purchase, and it measures net external cash rather than capital employed.
Realised vs unrealised
There is no equity curve, no deposit/withdrawal ledger and no current mark. positions.json is IBKR-only and stale at 2026-03-02 (171 days). This mostly does not matter, because capital fell £288k (Apr) → £123k (May) → £24k (Jun) and only £22,980 of cost basis was still open at window end — he ended flat, so realised P&L is ~99% of the true result. But the final 71 days of the window are empty and no post-2026-06-10 activity exists to check.
Missing cost basis
37% of all-time realised P&L (£54,590, 47 IBKR episodes) has no recorded purchase. In-window, BOT contributes £14,369 (10.9%) on zero recorded cost and is unverifiable; 20 episodes with uncosted sales carry £49,217 of imputed missing cost. Ex-BOT the window is £117,755 (+83.5%) — still ahead of QQQ, with the asterisk.
FX is fake
Every USD row converts at a hardcoded 0.773 for two years. These are USD returns relabelled GBP: ratios are comparable to USD SPY/QQQ, absolute £ figures are not what settled in his account.
Trading costs are 5–20× what the file shows
Recorded commission is £2,064.82 = 1.53bps of £13.5M turnover, 100% IBKR. 765 of 2,053 rows report zero commission. At 25–100bps of spread and FX markup on £3.32M of Revolut/T212 turnover that is £8k–£33k, i.e. 6–25% of net P&L. One Revolut fee statement would close this.
Revolut's per-row P&L is not time-consistent
Symbol totals reconcile; within-symbol attribution does not (a 2025-06-11 CRCL sell implies a $180.50 cost basis on a stock whose highest print to that date was $138.57). 29 HFG rows carry null price and null proceeds and are excluded from all entry-timing measures.
Effective sample size is closer to 7 than to 121
The window contained exactly one significant drawdown (2025-02-19 → 2025-04-08). He bought £196,892 against £36,605 sold — net £160,287 across 69 fills — and that is the strongest evidence of skill in the record. It is also n = 1.
Kept separate, never blended
CFDs (1,163 rows) are excluded from every figure on this page — a mistaken instrument the owner has retired, not part of the stock record and not netted against it. Options (94 rows, IBKR, six weeks of short-dated call writing against inventory he already held) are likewise reported on their own line and never folded into a stock number. Every return, drawdown and benchmark figure here is the stock book only.