What to expect
The pre-registered reference behind the expectation band, frozen 28 July 2026, nine days before the first trade.
The reference is sized for the account the validation runs on. At that size the system holds only a fraction of its roughly one hundred markets at any time, so this is the expectation for this account, not for the strategy in the abstract.
The expectation is a distribution, not a number
How to read: half of all half-years fall between the two dots, nine in ten inside the shaded band; the dotted line is zero.
What this shows. I cut the eighteen-year reference into every six-month stretch it contains, — overlapping windows of 126 trading days, and computed the Sharpe ratio and the return of each one. The chart and the table show how those half-years are spread out: the median half-year, the typical range, and the tails.
Why six months. That is the review rhythm of the validation phase. Every half year I take the realised window from the live track and place it in this distribution. The performance chart draws the 5 % and 95 % lines of the Sharpe row as the expectation band, so the comparison is visible every day, not only at review time.
Why a distribution and not an average. Because the average hides how much half-years differ even in a backtest that works over the long run: a quarter of all half-years lost money, one in twenty had a Sharpe below —, one in twenty above —. A bad half-year in the live track is therefore not evidence against the system, and a good one is not evidence for it. What would be notable is a window outside the 5 to 95 percent band; and even that is a reason to look, not a verdict.
| Rolling 6 months | min | 5 % | 25 % | median | 75 % | 95 % | max |
|---|
The 5 %, median and 95 % values of the Sharpe row are the expectation band drawn on the performance chart (−1.10 / 0.53 / +2.78).
Sharpe of the total system is in excess of the risk-free rate; the trend sleeve uses the raw daily Sharpe. Rolling windows are annualised with 256 trading days, following the research code.
Reference figures
Total system
USD May 2008 to May 2026 · in-sample referenceTrend sleeve
USD May 2008 to June 2026 · in-sample referenceSharpe of the total system is in excess of the risk-free rate, annualised with 252 trading days; the trend sleeve uses the raw daily Sharpe, annualised with 256 trading days. CHF: same trades, translated at daily USDCHF with a Swiss cash rate (xSARON) in the cash leg.
Backtest curve
| Year | Total |
|---|
* partial year. No benchmark and no band in this chart: the reference is the object being tested, not a race.
Why live will probably be worse
It is in-sample. The family weights of the system were set with the whole 2008 to 2026 history in view. There is no out-of-sample wall for the weight vector itself. An internal leave-one-out check puts the flattery at roughly a tenth of a Sharpe point; treat the reference as an upper bound, not a forecast.
Costs are modelled, slippage is not. The backtest charges commissions, spreads and roll costs per contract. It does not know about queue position, partial fills or the days when the market moves while an order is working. The validation phase exists to measure exactly that.
Paper fills. During validation, orders go to a paper account. Fills are simulated against real quotes, which is optimistic in fast markets and irrelevant in quiet ones. Nothing about live results can be inferred from paper fills; the point of the phase is process, not profit.
Two data worlds. The reference was computed on a research price history; the live system trades on the broker feed. In a check on 14 August 2026, 97 of 100 signals agreed between the two; the remaining differences were rounding and listing lags, but they exist.
Regime dependence. A few years carry a large share of the result; the yearly table shows which. If the coming years look like the quiet ones, the expectation band above says how that will feel: long stretches inside the lower half are normal, not a signal.
How long until we know, and what happens in the meantime
The first question most readers have is whether the system works. This is the honest answer to how long that question stays open.
Exact values: — / — years at two standard errors.
A Sharpe ratio measured over T years has a standard error of roughly the square root of (1 + SR²/2) divided by T. To separate the reference Sharpe of 0.66 from zero at two standard errors takes about a decade. To separate it from half that value takes several decades. So the validation phase cannot prove that the system has an edge, and I will not claim that it does. What it can prove is that the process holds: that positions match the model every night, that execution costs are what the backtest assumed, that the published numbers are the ones the account shows. That is why the question on the performance page is “is the track inside the band?”, not “does it beat the backtest?”.
What follows from this. The decision to start trading live, planned for late 2026, will not rest on the Sharpe ratio of the validation phase; it rests on whether execution, costs and reconciliation behave as the backtest assumed. Beyond that, the validation protocol schedules a written half-yearly review, and this distribution is its yardstick: the realised window is placed in it, and a window outside the band is a question the review has to answer. The band is a lens, not a gate; nothing switches automatically. What this page asks of you as a reader is the same patience it asks of me: judge the process now, and the edge in a decade.
What is deliberately not here
No instruments, signals, parameters or weights, and no variant runs. This page documents what I committed to expecting before the validation phase began; it does not document how the system works. The research log describes the method; the performance page shows what happens.
The reference is a single simulation of the frozen configuration over the period May 2008 to June 2026 on a research price history, sized for the validation account. The total system combines the futures programme with the passive ETF overlay at half of account value, both on the full account, and runs from May 2008 to May 2026 (end of the overlay data). Returns are compounded daily; drawdowns are computed on compounded equity. Sharpe ratios of the total system are in excess of the risk-free rate and annualised with 252 trading days, following the composition tool they come from; the trend sleeve and the rolling windows use the raw daily Sharpe and 256 trading days, following the research code. The CHF track uses the same trades translated at the daily USDCHF rate, with a Swiss cash rate in the cash leg. Rolling windows are 126 trading days; quantiles are empirical. The data file behind this page was built once from the frozen run and is not regenerated; its checksum is shown above, and any later change appears as a dated revision.
Data file SHA-256: 3f50c58ea1147aeb6e59bec91e40bde5245815d2783876c9f0a41c02d9950181.
Simulated performance. The figures on this page are hypothetical results of a backtest computed with the benefit of hindsight; they include modelled transaction costs but no live execution, and they are not the results of any actual trading. Simulated results have inherent limitations and do not represent, and are not indicative of, future results.
This is a personal research project, traded with my own capital. Nothing on this site is investment advice, an offer, or a solicitation; no client relationship is created. Figures from the validation phase come from a paper account with simulated execution. Futures trading carries a substantial risk of loss. Content may contain errors and can change without notice. Past performance is not indicative of future results.