Research
How Shock Engine is tested
Each test has a pass/fail criterion set in advance, and every change to the rules is disclosed with its selection risk.
Method
Principles
Parameters fixed once
The parameters were selected once, on Bitcoin, and are never re-estimated. Two entry conditions were added in October 2026 after studying the historical data (new positions in the calm regime only; shorts only in a bearish daily trend): the history is in-sample for them, and their selection risk is measured.
Criteria first
Each study states its criteria before the result is seen: profit factor, Sharpe, placebo rank, delay tolerance, event-study significance.
Placebos
Random entries with identical exits separate the value of the signal from the market’s drift and from the shape of the exits.
Across time and markets
Walk-forward re-selection on past data only, and transfer of the Bitcoin rules to Ethereum without recalibration.
Stress
Commissions doubled, entries delayed, every parameter perturbed at once: a real edge should bend, not break.
Reproducible
Each published result is tied to a code commit, data hashes and a parameter hash, with fixed random seeds. The same command gives the same files.
Evidence
Studies and verdicts
Single-market Sharpe ratios use 15-minute returns over each market's full test period; the portfolio uses daily returns. All figures are historical simulations after a commission of 0.045 % per order.
BTC/ETH portfolio, 50/50
2018–2026Sharpe 2.02, against 1.46 for BTC and 1.93 for ETH alone; strategy correlation 0.36; max drawdown −17.8%.
completeAfter buy & hold and six standard trend strategies are removed, alpha of 44.8% a year (t = 4.94); 81% of the average return is not explained by them. In-sample.
demonstrated · in-sampleBTC: event study, random entries, delayed entry
2017–2026Continuation after shocks confirmed; beats 100% of random-entry runs; a one-bar delay keeps 97% of the average trade. Sharpe 1.59 (15-min), 396 trades.
passBTC: walk-forward, 36-month calibration, 3-month tests
2020–202628 windows, settings re-selected on past data only: Sharpe 0.83, CAGR 14.3 %, max drawdown −27.6 %. Positive out of sample, clearly below the fixed preset.
positive, weakerBTC: walk-forward with longer windows (12 designs)
2020–2026Designs fixed in advance, after the result above: calibration of 36, 48 or 60 months or an expanding window from 2015, tests of 3, 6 or 12 months, same test period. Median Sharpe 0.97 (range 0.83 to 1.36); all 12 below the fixed preset (1.55). Window length does not explain the gap.
same conclusionBTC: parameter neighborhood (1,810 runs)
2017–2026All parameters perturbed together by up to ±20 %: median Sharpe 1.35 against 1.59; 100% profitable. The preset tops its neighborhood, as expected of rules selected on this data.
pass, selection biasETH: transfer of the BTC rules
2018–2026No ETH calibration of the parameters (the two entry conditions were set with knowledge of ETH data). Every pre-set criterion met: continuation, random entries (100%), delay (94%). Sharpe 2.05, 358 trades.
passETH: independent price feed (Dukascopy)
2019–2026Same verdict on a different data source: Sharpe 1.82, 206 trades.
passETH: parameter neighborhood (1,810 runs)
2018–2026At ±20 %, median Sharpe 1.94 against 2.05; 100% profitable. The rules sit mid-neighborhood: a plateau, not a peak.
passRule selection risk (CSCV)
2018–2026The 16 rule variants actually tried on the portfolio, compared over 12,870 splits of the history: the variant that is best on one half ranks above the median on the other half in 81% of splits (probability of backtest overfitting 19%). The gain of the calm-regime restriction over the previous rules is consistent over time but not significant once the number of variants tried is accounted for (deflated Sharpe 18%).
low overfitting, gain not provenExecution costs and capacity
2020–2026All 1,150 executions since 2020 were replayed on real Binance perpetual-futures trades. With the bot as now coded, execution costs average about 2.3 bps per execution at $1M of capital, for a net Sharpe of 1.92 over 2020–2026 (resampled 90 %: 1.40 to 2.41, funding excluded), still 1.83 at $5M, the largest capital tested, in the central estimate. Capacity on Hyperliquid, the intended venue, and our own market impact are not yet measured.
Binance measuredResidual alpha
Is it just trend following?
Could simple, well-known trend strategies reproduce the result? The daily returns of the portfolio are regressed on buy & hold BTC and ETH and on six standard trend strategies whose parameters were fixed before the study: momentum over 30, 90 and 180 days, a 55/20-day breakout, a 20/100-day moving-average crossover and an intraday breakout on 15-minute bars. Same bars, same commission, positions decided at the close and applied to the next day.
| Factor | Beta | t |
|---|---|---|
| Buy & hold BTC | −0.03 | −1.41 |
| Buy & hold ETH | 0.07 | 2.81 |
| Momentum, 30 days | 0.07 | 2.52 |
| Momentum, 90 days | 0.04 | 1.18 |
| Momentum, 180 days | −0.07 | −2.44 |
| Breakout, 55/20 days | 0.05 | 1.22 |
| Moving averages, 20/100 days | 0.04 | 1.36 |
| Intraday breakout, 24 h / 12 h | 0.15 | 5.59 |
| Benchmark alone | Sharpe | Correlation with Shock Engine |
|---|---|---|
| Buy & hold BTC/ETH | 0.82 | 0.09 |
| Momentum, 30 days | 0.96 | 0.34 |
| Momentum, 90 days | 0.45 | 0.20 |
| Momentum, 180 days | 0.36 | 0.08 |
| Breakout, 55/20 days | 0.84 | 0.33 |
| Moving averages, 20/100 days | 0.64 | 0.20 |
| Intraday breakout, 24 h / 12 h | 0.02 | 0.36 |
| Variant (fixed in advance) | Alpha per year | t |
|---|---|---|
| Main model | 44.8% | 4.94 |
| 19 trend strategies at once (favors the benchmarks) | 44.3% | 5.07 |
| First half of the period | 48.6% | 3.39 |
| Second half of the period | 44.1% | 4.41 |
| Weekly returns | 45.7% | 4.70 |
| Benchmarks without commissions | 40.2% | 4.62 |
Pre-registered before any computation: alpha counts as demonstrated only if t ≥ 3 and fewer than 1 % of month-block resampled histories give alpha ≤ 0. In-sample: the rules were chosen on data covering this period. The test shows that simple trend strategies do not reproduce the historical result; it does not show that the alpha will persist. Closest factor: Intraday breakout, 24 h / 12 h, which barely breaks even on its own after costs (Sharpe 0.02).
Short entries
Could the trend condition be luck?
Short positions are only opened when the daily trend regime is bearish. Two placebo tests, specified before they were run, ask whether a condition of the same size, chosen at random, would have done as well. They were run before the calm-regime restriction; every short entry comes from the calm regime under both rule sets, but the portfolio-level figure below includes the agitated-regime longs of the earlier rules.
The same number of short trades, kept at random, 50,000 times by market and year. The average risk-adjusted gain of the shorts actually kept beats 96.8 % of the draws (p = 0.032); for the portfolio Sharpe, 94.5 % (p = 0.055).
The bearish periods shifted in time, keeping their length and share of time. The gap between shorts in and out of the bearish regime is larger with the real dates than with 99.6 % of the shifts.
Both tests use historical data: they measure how unusual the result is, not how it will behave in the future. That question is left to the forward validation below.
Forward validation
A test that cannot be tuned
The plan was committed before the first observation: hypothesis, statistic, sample sizes, decision thresholds and stopping date. Nothing in it can change once it has started.
- Hypothesisshorts taken in a bearish daily regime earn more, risk-adjusted, than shorts taken outside it
- Observation starts2026-10-12
- Interim looksafter 25, 50 and 75 closed shorts in the bearish regime, at p < 0.001
- Final lookwhen both groups reach 100 closed shorts, at p < 0.047; at the latest 2031-10-01
About 89 trades a year, few of them shorts
Reaching 100 closed shorts in each group should take years rather than months, less in a falling market. The plan sets a monthly evaluation and fixed looks; a result is read only at those looks.
History
Earlier work
Shock Engine started as a TradingView script whose original settings lost money on Bitcoin after costs. The first research passes diagnosed why, showed that re-optimizing many parameters chases noise, and that among regime-based selections only a calm/agitated volatility split was worth pursuing: it became the basis of the current rules. In October 2026 the rules were narrowed to the calm regime: new positions are no longer opened in the agitated regime.
Reproducibility
Current version
| Research version | shock-engine-2026.10.13 |
|---|---|
| Code commit | 997435b976491bffcc356ce8b3e916cbcd02a17b |
| Parameter hash (SHA-256) | c6ac30464bce14ddb6e5625cbcf08f68e69e3cc5e5454a0c2268583070a74435 |
| Consistency checks | 40 of 40 pass |
| Manifest | shock-engine-manifest.json |
| Research paper | Shock Engine research paper (PDF) |
| Institutional dossier | Strategy, execution and capacity (PDF) |
| Full report | BTC/ETH portfolio report (HTML) |
Important information. All performance figures on this site are hypothetical: they come from a historical simulation of fixed rules, after modeled commissions, and do not represent actual trading. Simulated results have inherent limitations. They benefit from hindsight in the design of the rules, they do not reflect slippage, funding, market impact or the liquidity of a real account, and past simulated performance does not predict future results.
Nothing on this site is investment advice, a recommendation, or an offer or solicitation to buy or sell any financial instrument or to provide investment management services.