Methodology · research receipts
How every review earns its numbers
Corrly's advisor reviews separate observed data, calculated estimates, and advisor judgment. Each delivered figure names its data window and estimator in an attached calculation summary. The validation record below covers the bounded allocation-sensitivity method; it is not a blanket promise that every historical relationship will persist.
Walk-forward testing
We never grade the model on data it trained on. The balanced allocation is computed from 378 days of history, then judged only on the next 63 days it never saw — and the window rolls forward and repeats, 405 times across 70 realistic retail portfolios. The volatility reduction we show holds up out-of-sample, and the gap between in-sample and out-of-sample results — the classic overfitting tell — is 0.6%.
Per-report provenance
The current sample review includes a calculation summary recording the data provider, retrieval time, market-data window, exact weights, methods, and a hash of the retrieved price table. It does not preserve the licensed raw price series and is not a reproducible frozen input snapshot. Privately retained snapshots and automated report-to-summary verification are required before any real-client delivery. You can inspect the calculation summary behind the sample review.
Did we just get lucky?
We compute the probability of backtest overfitting (PBO, Bailey & López de Prado) — roughly, the chance that the method that looked best in testing was just the luckiest of the ones we tried. For the volatility claim it’s 0.17, comfortably below the 0.5 coin-flip line.
What we deliberately don’t claim
The same test says ranking methods by risk-adjusted return (Sharpe) is indistinguishable from luck (PBO 0.51). So Corrly never claims the balanced version earns you more — only that it has historically swung less. We also don’t predict returns, forecast events, or react to news: the model describes how your holdings have moved, it doesn’t guess where they’re going.
The machinery, briefly
The current sample review uses 504 trading days of daily simple returns and ordinary sample covariance for its variance, contribution, correlation, volatility, and Gaussian-reference figures. Its allocation-sensitivity comparison minimizes in-sample volatility over the same holdings and is labeled as hindsight, not as a recommended portfolio or forecast. The separate walk-forward experiment summarized above evaluated an exponentially weighted, bounded allocation method on unseen windows. Corrly does not use machine-learned return predictions.
Educational and informational only — not investment advice. Past performance does not predict future results. Numbers on this page come from logged validation runs and update when the methodology does.