A five-minute Bitcoin scalping setup, before and after trading costs

What happened in one sentence

The frozen 2025 evaluation of one mechanical MoneyZG five-minute Bitcoin scalping setup recorded 1,833 completed trades, gross expectancy of +0.0083R per trade and net expectancy of −0.6632R after modeled fees, slippage and funding. Its two-sided 95% confidence interval for mean net R was [−0.7832, −0.5433]R, entirely below zero. The study closed COMPLETE with the scientific classification NEGATIVE_EXPECTANCY_SUPPORTED.

The original video is not a documented trade ledger. This experiment says something about my specified interpretation, with specified costs; it cannot tell us whether the presenter traded different discretionary variations successfully.

The strategy on the screen

In MoneyZG’s May 2021 video, the indicators are familiar: a 20-period EMA, a 50-period EMA, and Williams Fractals set to Periods 2, all on a five-minute chart. The video demonstrates setups on a chart displaying BTCUSDT, but it does not establish the precise contract or price feed used here.

When the 20 EMA is above the 50, the setup looks for a confirmed down fractal: a local price low between the averages. That is a candidate long. When the faster EMA is below the slower one, a local up fractal between the averages becomes a candidate short. The source discusses a stop at the 50 EMA and targets as risk multiples, with some discretion. It also shows optional position additions.

An up fractal here names a local high, not a bullish instruction. The marker is drawn on the historical pivot candle. The trading decision cannot be made at that pivot without seeing the two subsequent candles.

A fractal needs two future candles to confirm

Williams Fractals (Periods 2) use a five-candle neighborhood. With a possible pivot at t, the model waits for the t+1 and t+2 candles to close, then allows entry no earlier than the open of t+3. The study required the fractal extreme to fall strictly between the pivot-time EMAs and the EMA trend ordering to hold at both pivot and confirmation.

This timing is important, but I did not infer that MoneyZG recommended look-ahead trading from the video’s retrospective chart examples. The source explicitly discusses waiting for confirmation. The question was how to turn those examples into a consistent rule that cannot use information from a future candle.

Other decisions needed equal care. The source’s target and position-management instructions do not uniquely specify a backtest. I therefore fixed a stop at the confirmation-time EMA50, a 1.5R target, no stop trailing, one position at a time and no pyramiding. Strict tie handling and EMA initialization were also specified.

These are study conventions, not claims that the video requires every trader to follow them.

One year, reserved before the data were acquired

The study used BTCUSDT USD-M perpetual futures five-minute candles and funding history. Its scientific holdout covers exactly 1 January 2025 00:00 UTC inclusive to 1 January 2026 00:00 UTC exclusive. Five hundred preceding bars provided indicator warm-up; warm-up pivots could not create eligible trades.

The 2024 development period was used to verify the evaluator and investigate inference. In that exploratory sample, 1,959 trades averaged −0.5773R net. The decision rules were not revised to improve that result.

After independent trade-level reconstruction, a statistical-method validation and a platform execution/verification test, the scientific contract was frozen on 29 October 2026. The 2025 sources were acquired only afterward. This creates a reserved, freeze-before-acquisition historical holdout; it does not turn historical market observations into live or genuinely forward-collected evidence.

All 105,120 expected five-minute evaluation bars were continuous, along with 500 preceding warm-up bars and 1,095 expected funding settlements. All frozen source-adequacy requirements passed. The records bind the acquired input bytes to the frozen specification and sealed result.

Costs relative to a short stop

In the simulated execution model, entry used the next open, with 5 basis points of taker fees and 2 basis points of adverse slippage per side. Stops and targets were tested against candle highs and lows, with conservative stop-first treatment if both were touched within a candle. Funding used recorded settlements and an adverse assumption when the available candles could not order funding against a possible intrabar exit.

The accounting used a mechanical one-BTC-per-trade position size. The mean modeled costs per completed trade were:

Slippage is already reflected in the simulated entry/exit prices and net R. It is shown separately to explain the difference, not subtracted again.

This is why the basis-point convention matters. A fee of several basis points can be a large fraction of R when the stop is near the entry. In the development diagnostics, the median stop distance was 31.5 basis points of the entry price. Those development-period distances help explain the cost sensitivity; I am not representing them as a separately calculated 2025 stop-distance statistic.

The model’s aggregate charges in notional USDT terms were 185,701.29 in fees, 74,280.52 in slippage and 38.35 in net funding across the simulated trades. These are not actual exchange debits, and the model does not establish what an account could have filled.

What the frozen decision rule actually required

The primary question was whether the model had positive mean net R per completed trade, established by a lower confidence limit above zero. Merely having a positive gross return, or a good-looking example, was insufficient.

The inference grouped 365 consecutive UTC days into five 73-day batches, pairing daily net-R totals with counts of completed trades and keeping zero-trade days. A ratio-influence standard error and Student-t interval with four degrees of freedom estimated uncertainty in the total-net-R / completed-trades ratio.

The procedure was chosen and validated on prespecified synthetic scenarios before the holdout was acquired. Its coverage on those scenarios was close to 95%; that is not a universal guarantee under structural breaks, very long memory or other untested dependencies.

The real holdout had 355 exit days, and the five batches contained 383, 371, 359, 346 and 374 trades. It passed the minimums of 200 trades, 30 exit days and at least one trade per batch. No position remained open at the final evaluation boundary.

The 95% interval, [−0.7832, −0.5433]R, is entirely negative. The fixed positive-expectancy test failed, and the recorded outcome supports negative modeled mean expectancy under the stated statistical assumptions.

What the model cannot see

The 1,833 simulated trades comprised 938 longs and 895 shorts. The net win rate was 35.19%; 1,164 trades stopped out and 669 reached target. Three stop/target candles were ambiguous, and ten trades had an intrabar funding ambiguity. The prescribed conservative rules resolved those cases.

Five-minute OHLCV, however, cannot reconstruct the actual path within a candle, the bid/ask spread, depth, queue priority, or fill slippage at an exchange. These mechanical outcomes cannot be presented as executed historical trades. The model also excludes discretionary filters and pyramiding that may change what a human would select.

The fixed 1.5R target is probably the most consequential simplification of the original tutorial. MoneyZG discusses targets and position management with some discretion, whereas this evaluation requires the same exit logic on every trade. By contrast, although the assumed 2 bps adverse slippage per side is uncertain, modeled taker fees alone were far larger than the gross mean return. The negative result should therefore be read as evidence about this specific mechanical implementation, not every possible discretionary variation.

What I can say is narrower, and more useful: in a complete frozen historical evaluation, this confirmation-first version did not earn enough gross modeled R to cover its stated trading costs. I have not established that every possible MoneyZG interpretation is unprofitable, or that the video’s creator reported a backtested return that this run disproves.

A review of someone else’s scalping backtest should therefore start with the entry time, initial stop distance, and cost expressed as a fraction of that distance—not just a win rate or a screenshot of favorable trades.

For a reusable set of questions, see Before you trust a backtest.

Evidence, independent verification and data attribution

The authoritative study lives in the separate MoneyZG research repository. Its sanitized methodological handoff, machine-readable result summary, and chart aggregates make the reported figures inspectable without redistributing raw market-data archives.

Frozen specification SHA-256: 14d326cf92bb92fefd5ef65fcaa6fe8e074aeea75aca3f9073df86ad1b44697c

Sealed result SHA-256: 4ead222258da57befb7f662230d24706c17177940f5484f0f81ced2c918eca84

Independent semantic verification SHA-256: c1c9806cc23b39485fb209bf7ec4d490fbf4470396f12285b3fd017316c25933

Run identity: 20261029T132412Z-59e969df48e1. Platform integrity and independent signal, ledger, funding and inference replay both passed. The descriptive-summary helper was shared with the main implementation rather than independently rewritten, so I do not claim that every reported field had a separate implementation.

Market-data attribution and reuse. Historical OHLCV inputs came from Binance Vision; funding rates and settlement marks came from Binance’s USD-M funding-history endpoint. Market records were transformed into hypothetical trades and aggregated research charts and statistics. Binance Vision Dataset Terms describe noncommercial research access and refer to CC BY-NC-SA 4.0, including attribution and share-alike for covered adaptations. The study’s selected derived chart data are presented for noncommercial educational use under those conditions where applicable; this is not a blanket relicensing of unrelated site content. No raw archives, funding payloads or trade ledger are hosted here, and Binance has not endorsed the analysis. The relevant terms for older archives and the keyless funding REST endpoint require interpretation; the study’s rights note documents those open questions rather than claiming individual permission.