00 / Performance
We would rather show you the method than a number we cannot defend.
Most platforms lead with a curve. We lead with how the curve is judged, because a curve is only as honest as the test behind it. This page shows exactly how we decide whether an edge is real, and reports the result in full: none of the current sleeves clears multiple-testing deflation, and the live paper record is still too short to overturn that result.
Published strategy record: paper only, no funded performance.
02 / What we measure
Six ways to catch ourselves out.
Search enough ideas and one will look brilliant by luck. These six tests separate that from a real edge. Each asks the same question a different way: would this survive if the world had not been so kind?
Purged walk-forward
Gate: Positive, stable out-of-sampleOut-of-sample by construction, with a purge and embargo around every train and test boundary. If an edge only exists in-sample, it disappears here.
Deflated Sharpe Ratio
Gate: 0.95 to deployA Sharpe discounted for the number of trials behind it and the non-normal shape of real returns. Given how hard we looked, how impressed should we be.
Probabilistic Sharpe Ratio
Gate: High confidence above zeroThe probability that the true Sharpe sits above zero, given the sample. A confidence statement, not a point estimate.
Probability of Backtest Overfitting
Gate: Below 0.20Estimated by combinatorially symmetric cross-validation. The chance the best in-sample configuration is no better than median out-of-sample. Lower is better.
Combinatorially-purged cross-validation
Gate: Consistent across foldsPerformance estimated across many purged train and test combinations, so the headline figure is not one lucky path through history.
Must-beat-baseline gate
Gate: Beats baseline by a marginEvery configuration tried is counted and fed into the deflation, and the strategy must beat a simple honest baseline by a margin. Beating zero is not the bar.
// rebalance: 72 bars (3 days) / purge: 72 bars / embargo: 168 bars// walk-forward: rolling, expanding train / trials: 8 / baseline: blend-only, same legs
02 / The gauntlet
One question, asked six ways.
The same field of signals, run through every test in turn, until one verdict remains or the edge fails.
-
01 / Purged walk-forward
Always out-of-sample.
The backbone. Train on the past, test on an unseen future, with a purge and embargo around every boundary. If an edge exists only in-sample, this is exactly where it disappears.
Gate / positive and stable out-of-sample
-
02 / Deflated Sharpe Ratio
Discounted for how hard we looked.
A Sharpe ratio discounted for the number of trials behind it and for the non-normal shape of real returns. Given how hard we searched, how impressed should we actually be.
Gate / above zero after deflation
-
03 / Probabilistic Sharpe Ratio
A confidence, not a point estimate.
The probability that the true Sharpe sits above a chosen threshold, given the sample we have. A statement of confidence rather than a single flattering number.
Gate / high confidence above threshold
-
04 / Probability of Backtest Overfitting
The signature of a curve-fit.
Estimated by combinatorially symmetric cross-validation: the chance the configuration that looked best in-sample is no better than the median out-of-sample. Lower is better. A high reading is the classic tell of a curve-fit.
Gate / low, well below 0.5
-
05 / Combinatorially-purged cross-validation
Not one lucky path through history.
Performance estimated across many purged train and test combinations, so the headline figure is the behaviour of the method across histories, not the single kindest route through the past.
Gate / consistent across folds
-
06 / Must-beat-baseline gate
Beating zero is not the bar.
Every configuration we tried is counted and fed into the deflation, and the strategy must beat a simple honest baseline by a meaningful margin. A result that only beats doing nothing has not earned a dollar of risk.
Gate / beats a simple honest baseline by a meaningful margin
03 / The facts
What is true today.
These are the only numbers on this page. Every one is a property of the engine and its data, not a performance claim. Hold us to them.
- 0 No look-ahead
- 1 Cost authority, research and paper
- 2020 Crypto history starts
// Performance is absent from this ledger on purpose. Canli Capital has not traded real capital. There is no live return to report, and a backtest is not one.
04 / Results
Here is the curve, and what it did not clear.
This is the validated out-of-sample result for AlphaForge: the equity curve, the Deflated and Probabilistic Sharpe ratios, the Probability of Backtest Overfitting, the trial count behind them, and the baseline it had to beat. It did not clear the bar, and we publish it in full.
This is a real, validated result. It did not clear purged walk-forward and the deflation, and we publish it in full. The crypto-only edge is a null, and the path forward is breadth.
Source: grand_backtest/20260616T143620Z / As of 2026-06-17 / Crypto perpetuals / Out-of-sample 2022 to 2026
05 / Capacity
Size has a price.
An edge that exists at one notional may not exist at ten. The capacity curve traces expected return net of cost against deployed capital, so the point where size starts to erode the edge through impact and funding is visible, not assumed. It draws from a validated artifact or it stays blank.
Annualized Sharpe versus deployed capital. The thin edge decays as size grows and turns negative well before institutional scale.
06 / Standing
No sleeve has cleared deflation. We are telling you on purpose.
AlphaMax, AlphaTrend, AlphaVintage, and AlphaForge all remain below the programme's multiple-testing gate after the full research search is counted. The current paper book tests diversification and implementation; it does not certify that any sleeve has alpha. So we report the null instead of dressing a backtest as a promise. Canli Capital does not claim to always win. It runs the test in the open and reports what passes and what does not.
We would rather post a null than a number we cannot defend.
The paper book combines three sleeves at equal thirds plus a separately disclosed strategic overlay. Their measured diversification is useful but smaller than we first claimed. Only a long, continuous, broker-reconciled forward record can turn that construction into evidence.
07 / The current book
What the sleeves are holding, right now.
The book runs three sleeves: AlphaMax (US-equity 12-1 momentum), AlphaTrend (managed-futures trend) and AlphaVintage (point-in-time macro surprise), combined into ALPHAC at equal thirds. AlphaForge (crypto funding carry) is suspended from the book since 2026-09-24 and is not shown until it is repaired and re-admitted. None clears multiple-testing deflation at the current research identity count, and the forward record is too short to estimate a forward Sharpe. Below are the current paper positions read from each dedicated Alpaca account.
Honest forward Sharpe not loaded full-book in-sample not loaded
Holding cash, no open positions
The funding-carry loop stands aside when carry does not clear costs; when it does, it trades a market-neutral book. The state above is read live from the record, in whichever state the sleeve is in.
The book holds AlphaMax (US equity momentum), AlphaTrend (managed-futures trend) and AlphaVintage (point-in-time macro surprise) at equal thirds. Their research curves are near-uncorrelated, and the average pairwise correlation is slightly POSITIVE (current measurement). We used to call that decorrelation “the edge” and we withdrew the claim: the diversification is real but smaller than we said. The disclosed overlay is commoditized beta, held as a separate labelled line, never blended into the neutral core.
// The three Alpaca sleeves above are current broker positions, not strategy targets. AlphaForge is suspended from the book while it is repaired and keeps paper-trading off the record. No real money is in play.
// Broker reconciliation evidence: not loaded / account scope unavailable / position count unavailable / timestamp unavailable.