Week one review: replay, costs and checkpoints
Eight days of the DEMO and PAPER accounts against the model replay: tracking gap, loss attribution, round trips, defects found and the day 30, 60 and 90 review lines.
Scope: the Binance DEMO and Hyperliquid PAPER accounts from their flat start on 27 September 2026 to the 4-hour close at 04:00 UTC on 5 October 2026, 8.2 days and 49 four-hour bars per account. Both accounts trade simulated capital. No rule or parameter was changed on the basis of this review; the checkpoints at the end were fixed before the next result is known.
Method
- Model replay. The deployed rule was re-run bar by bar on the same candles from the same flat start, net of 0.07% per unit traded and Binance's funding, and compared with each account at every 4-hour close and leg by leg.
- Loss attribution. The Hyperliquid account's result was split into the model's gross market P&L, the path difference to the model, execution against the close of the decision candle, fees and funding.
- Historical context. The same rule was run from a flat start on every day since 1 January 2021, 2,096 overlapping windows, to place an 8-day result in its distribution. The same distribution, recomputed from the site's research data, is now drawn as a 5th to 95th percentile band behind each account's chart on the track record.
- Operations. Cycle logs, restarts, stop orders and alerts on both accounts.
Results
Tracking against the model
| At the 5 Oct 04:00 UTC close | Binance DEMO | Hyperliquid PAPER |
|---|---|---|
| Return since start, account | −1.11% | −1.12% |
| Return since start, model replay | −1.08% | −1.07% |
| Account minus model | −0.03 pp | −0.05 pp |
| Correlation of 4-hour returns with the model | 0.996 | 0.994 |
46 of 48 live legs had the same direction as the model and a size within 15% of it. The two exceptions cleared the resize threshold by less than 0.3 bps of equity, inside the replay's approximation error. Turnover was 1.71 times equity against 1.70 in the model. Each account ran 49 of 49 cycles with none missed or repeated, every open position had a stop about 25% from its entry price (an exchange order on Binance, simulated on the paper account), and no risk limit came near its threshold.
Loss attribution
| Hyperliquid PAPER, % of start equity | Contribution |
|---|---|
| Model gross market P&L | −0.93% |
| Path difference to the model | +0.05% |
| Execution against the decision close | −0.10% |
| Fees | −0.08% |
| Funding | −0.07% |
| Total (components rounded) | −1.12% |
About 80% of the loss is the rule's gross market loss and about 20% is costs. In the model the loss came from ETH (−0.40% of equity), LINK (−0.33%) and BNB (−0.28%); ETH and BNB rose over the window, but the positions were built near the 27 September highs and reduced lower. BTC rose 1.3% within a 5.8% low-to-high range, and its 8-day realised volatility was at the 8th percentile of 2021 to 2026.
The 8-day result in the backtest distribution
| Return 8.2 days after a flat start, 2,096 windows | Value |
|---|---|
| Median | −0.18% |
| Share of windows below the start | 53.9% |
| 5th / 25th / 95th percentile | −3.57% / −1.41% / +8.25% |
| Median in quiet-market windows | −0.41% |
| Percentile of −1.26% and −1.29%, the two accounts' marks when the analysis began | 28th |
Both accounts stayed inside the 5th to 95th percentile range at all 49 bars. An 8-day sample has a standard deviation of about 3.6% at the backtest's volatility, against an expected return of about +0.8%, so the result is about half a standard deviation below expectation and carries no information about the strategy's edge.
Costs
| Cost per unit traded: fees plus execution against the decision close | bps |
|---|---|
| Backtest base case | 7 |
| Hyperliquid PAPER, 23 fills | 10.5 |
| Binance DEMO, 23 fills | 11.5–12.3 |
The excess over the base case is about two standard errors and amounts to 0.06–0.08% of equity over the 8 days. Neither venue measures production execution: several demo taker fills landed about 10 bps through the quote and outside the production market's 1-minute range at the time, which points to the demo order book, and every paper fill is charged as a taker. At 10.5 and 12.3 bps the backtest returns 39.6% and 38.5% a year (Sharpe 1.49 and 1.46) against 41.8% (1.56) at the base case, all net of funding. The base case stays at 7 bps until production fills exist. Hyperliquid funding was 0.069% of equity, 2.07 times what the same positions would have paid at Binance rates; 8 days is too short to act on.
Exits, trims and round trips
| Binance DEMO | Hyperliquid PAPER | |
|---|---|---|
| Fills | 23 | 23 (16 in the published log) |
| Exits including partial reductions | 11 (9 trims, 2 full closes) | 11 (8 in the published log) |
| Exits in profit | 0 | 0 |
| Completed round trips | 2 (BNB, ETH) | 2 (BNB, ETH) |
| Round trips in profit | 0 | 0 |
Counting every partial reduction as an exit gives 0 of 11 per account, but the 11 exits come from five positions per account and are not independent outcomes. The site now reports completed round trips, each position from its opening fill to its final exit with its trims credited to it, and keeps the exit count as a secondary figure. In the backtest, 28% of 8-day windows with two or more completed round trips had none in profit, so 0 of 2 says little. None of the 15 logged trims across both accounts was profitable. In a bar-by-bar backtest re-run, 76.8% of trims were profitable, and about 70% in quiet windows. That re-run trims more often than the bot, so the comparison is indicative only; a separate check of the trims is open.
Defects
None of the defects below changed a position or a return. They affect records, reporting or behaviour on paths the accounts have not yet taken.
| Defect | Effect | Status |
|---|---|---|
| The Hyperliquid fill log is missing the first 7 fills (27–28 Sep). | The published record shows 16 fills and 8 exits instead of 23 and 11. Equity and returns are unaffected. | Backfilled on 7 Oct from the paper ledger and event log with the bot stopped; the record now shows every fill. |
| The Binance demo account read its signal candles from the demo exchange, not production Binance. | Demo closes differed by 2.8 bps on average (up to 77 bps) and caused two extra SOL resizes. | Fixed and deployed on 7 Oct after code review. Every venue will read production Binance candles, as the backtest does. |
| Three different "since start" returns were published for the demo account (−0.84%, −1.10%, −1.30%). | Each was correct at its own time, but none carried a time stamp. | Site fixed: one basis, equity at the latest snapshot against start equity, shown with its time. |
| On 6 Oct the Hyperliquid paper account closed LINK with a 39.8 order against a position of 39.899999999999 that its fills summed to in floating point, leaving 0.1 LINK. | The remainder rounded to zero on every later cycle and was skipped as below the minimum order size, so the account showed five positions. Value $1.35. | Fixed and deployed on 7 Oct: quantities within a billionth of a step snap to the step before rounding, and the remainder is closed. |
| The weekly report and dashboard computed a Sharpe ratio from 7–8 daily returns, one series mixing in a 4-hour return. | A ratio from so few returns is noise. | No ratio from fewer than 30 days of live data is shown on this site. |
| A Binance timeout reply (HTTP 408, code −1007, "execution status unknown") was treated as a rejection. | An order that reached the exchange could go unlogged or rest unnoticed. Seen once, on a read request, with no effect. | Fixed and deployed on 7 Oct after code review. The reply is treated as an unknown outcome and reconciled by the order's client id. |
| No orderly shutdown; on 1 Oct the server's automatic package update restarted the services at 04:40 UTC. | A post-only order resting during a restart could have been left on the book. | Fixed and deployed on 7 Oct after code review. Shutdown finishes or cancels the order in flight, and orders left by a previous run are cancelled at start. Moving the update timer away from the 4-hour closes is an operator step. |
| Binance fees and funding on the demo account were estimated, not measured. | Cost figures for the demo account carried an error of about 0.03% of equity. | Fixed and deployed on 7 Oct after code review. Fees and funding are read from the account's income history each cycle. |
| Execution was measured only against the quote at order time. | Not comparable with the backtest, which prices fills at the decision close. | Fixed and deployed on 7 Oct after code review. Every fill stores its shortfall against the decision candle's close. |
| The exit win rate counted partial reductions as exits. | A headline of 0 of 11 mixed correlated partial reductions with completed trades. | Site fixed: completed round trips with their count. The bot's own reports followed on 7 Oct. |
| Risk-gate decisions were overwritten each cycle, and the weekly report's cross-venue line never printed. | A gate that fired would have left no record; the venue comparison was missing. | Fixed and deployed on 7 Oct after code review. Gate inputs and decisions are saved with each cycle, and the comparison line prints. |
| The PAPER description did not say that all paper fills are charged as taker. | Readers could assume maker fees were simulated. | Site fixed. |
All code fixes above went live on 7 October 2026 at 06:33 UTC, outside the 4-hour decision window, after an independent code review. Moving the server's update timer remains an operator step.
Review thresholds
Basis: per account, return since the flat start on 27 September 2026, measured at the last closed 4-hour candle. Reference distribution: the 2,096 flat-start backtest windows above, net of 0.07% per unit traded and Binance funding. These lines were fixed on 5 October 2026 and will not be moved after the fact; the tracking lines are recalibrated once, at day 30, and then frozen.
| Day 30 (27 Oct) | Day 60 (26 Nov) | Day 90 (26 Dec) | |
|---|---|---|---|
| Backtest 5th / 50th / 95th percentile | −6.6% / +1.3% / +19.0% | −9.3% / +4.6% / +26.4% | −8.5% / +7.1% / +31.0% |
| Share of backtest windows below the start | 41.8% | 30.5% | 24.6% |
| Watch: noted in the weekly report, no action | below −2.5% | — | — |
| Review: audit execution and data against the model replay; no parameter change on this basis alone | below −6.6% | below −9.3% | below −9.3% |
| Hard review: consider pausing new entries while a walk-forward re-check runs | below −9.1% | below −12.8% | below −13.9% |
| Account minus model replay: review if worse than | −1.5 pp | −2.2 pp | −2.8 pp |
| DEMO minus PAPER: review if the gap exceeds | 1.5 pp | 2.5 pp | 3.0 pp |
Given the first week, the probability of being below the start is 61%, 42% and 35% at days 30, 60 and 90, against 41.8%, 30.5% and 24.6% unconditionally. The day-90 review line is set at −9.3% rather than the 5th percentile of −8.5% so that it never tightens between checkpoints.
Lines that apply at any time:
- Drawdown: review at worse than −15% within the first 90 days; hard review at worse than the backtest's maximum of −18.8%. The bot's 25% drawdown circuit breaker remains the last line.
- Tracking: review if the 30-day correlation of 4-hour returns with the model replay falls below 0.97, if the two accounts' correlation falls below 0.98, or if a weekly replay finds a live leg the rule does not produce.
- Execution: review if fees plus shortfall exceed 15 bps per unit traded over at least 100 fills (backtest Sharpe about 1.39 at that cost); hard review above 30 bps (about 1.11). Review if 30-day turnover exceeds 1.3 times the model's.
- Funding: review if Hyperliquid funding exceeds twice the Binance-rate equivalent over 60 days.
These lines are built to detect a broken system and cannot confirm the edge. A strategy with no edge and the same volatility would breach the day-90 review line only 18–23% of the time, and separating a Sharpe ratio of 1.55 from zero at two standard errors takes about 1.7 years of live data. The backtest distribution comes from a rule selected on the same history and should be read as optimistic.