Factor-neutral alpha, tested where it counts
Every figure on this page was produced by the training pipeline on data the model never saw, is charged with realised trading costs, and is published unchanged. Nothing here is fitted after the fact.
Factor-neutral long/short book, net of costs, out of sample
This is the portfolio a manager running the system would actually hold: a broadly diversified long/short book across the liquid US universe, rebuilt at every rebalance date from the model's forecasts net of each stock's own trading cost, with its exposure to every risk axis listed further down solved to exactly zero and a hard cap per position. It is dollar-neutral, market-neutral and style-neutral by construction. The sizing and construction rules are proprietary.
Equity curve and drawdown at 1× leverage
Net of realised round-trip costs, before financing. Each rebalance date accrues its share of the holding-period return, so overlapping tranches are neither double-counted nor dropped.
Months are the sum of the accrued per-rebalance returns falling in that month. The final month is partial.
Walk-forward: refitted every year, built by the production optimizer
The deployed configuration is refitted at annual origins on data before the origin only; every monthly rebalance is scored with the model that was live at the time, and the portfolio is built by the same optimizer users run on this platform: cost gate, net-of-cost objective, factor-neutrality band, industry cap, beta target, position bounds and a volatility target, on a covariance of the previous year of daily returns as it stood on that day.
The deployed configuration is refitted at annual origins on data before the origin only; every monthly rebalance is scored with the model that was live at the time and built as the factor-neutral, cost-gated long/short book: exact neutrality on every risk axis, dollar-neutral, beta 0, positions sized by residual risk at a few basis points each across the whole liquid universe, at a fixed gross exposure. Feature sets and hyper-parameters were selected once on the full history; only the model fits are walk-forward.
Both books use identical predictions and dates. The closed-form book is the training metric's own construction; the optimizer book is what the platform builds for a mandate with these settings. Neither includes financing or size-dependent market impact, which the capacity study adds.
Net of the realised per-name spread on the notional traded; financing, borrow and size-dependent market impact are added by the capacity study. The S&P 500 is measured over the identical rebalance dates and holding horizon.
Why the book is factor-neutral
Intratio does not run a factor strategy. The model's return is measured, and delivered, with every common risk premium removed. Here is why that choice is deliberate.
1You already own the factors
Market beta, size, value, momentum, quality, low volatility, rate and credit sensitivity can all be bought for a few basis points through index and smart-beta products. A manager who charges alpha fees for returns a factor ETF would have delivered is mis-selling. With every one of those exposures set to zero, each basis point in the curve above had to come from stock-specific insight.
2Factor returns are cyclical and crowded
The large quantitative drawdowns of the last two decades — August 2007, the 2018–2020 value winter, the 2020–2021 momentum reversals — were factor events shared across managers. A book with no exposure on those axes is not hostage to a regime, and it does not crowd with the rest of a family office's manager line-up. It diversifies by construction, not by promise.
3It makes the track record attributable
When exposures are zero at every rebalance, the return cannot be explained by a bet on the market or on a style, so the significance test below measures skill rather than the luck of the cycle. It is also why the raw signal's decile spread further down is larger than the book's return: part of that spread is factor exposure we choose not to carry.
4It disciplines the model
The model is trained to explain the part of returns that factors cannot. That objective removes the easiest and most crowded sources of apparent predictive power and forces the learning onto information that is genuinely stock-specific — the only kind that survives when the style cycle turns.
The method used to neutralise, the portfolio-construction rules, the feature set and the model configuration are proprietary and are not disclosed on this page. The statistics that prove the property are.
Mean standardised exposure of each score decile of the raw model output, out of sample. A naive top-minus-bottom portfolio would silently buy high-beta, high-volatility, smaller, momentum-rich names against large, low-beta, value-rich ones.
Max |exposure| is measured at every rebalance on the solved book. The ex-post t-statistic regresses the book's realised return on each axis's realised factor return over the test window: |t| below 2 means no detectable co-movement after the fact. Exposures are fixed at formation; over long holding periods they can drift before the next rebalance, which is what a non-zero ex-post t would reveal.
The raw ranking: top decile against bottom decile
Equal-weight long the top 10% of scores and short the bottom 10% at each rebalance, each leg 100% of equity. This is a diagnostic of the ranking, not the investable book: it carries the factor tilts shown above and no position sizing.
Calibration by score ventile and stability of the rank signal
The model is built to be right at the extremes. It is not asked to rank the middle of the universe, and the ventile table shows exactly that: flat in the centre, steep and statistically significant in the tails.
What the same book delivers at different gross exposure
The book is leverage-invariant before financing: return, volatility and drawdown scale together. Financing is charged on gross exposure in excess of equity at the spread over the reference rate, which cancels for a dollar-neutral book.
How the numbers were produced
The test window is strictly posterior to everything the model was fitted on, with an embargo in between so that no forward return overlapping the training period leaks into the test.
- Plateau, not peak. Model capacity is chosen from the stable plateau of the held-out validation curve rather than at its noisy maximum, and that choice is vetoed whenever it disagrees with cross-validation.
- Robust ranking of candidates. Candidate configurations are compared on a paired, autocorrelation-robust lower confidence bound of the neutral spread, never on the best-looking point estimate. The winner is the configuration that is hardest to beat, not the luckiest.
- Ensemble of independent fits. The test-phase model averages independently seeded fits so that no single random draw drives the result.
- One shot at the test window. The test window is opened once, after every modelling choice is frozen on training and validation data. Nothing is tuned against it.
- In-sample fits. The production model is refitted on the full history before deployment. Its in-sample statistics are spectacular and meaningless; they are excluded here and used only as a sanity check that the refit worked.
- Cherry-picked windows. The test window is the single most recent block of history the model was not trained on. There is one test window per training run, shown in full.
- Untested claims. Every statistic on this page is read from the artifact files written by the training run, on one shared accrual and annualisation convention. The page does no fitting of its own.
Limitations and how to read the statistics
Every out-of-sample prediction, one point each
Model score against realised return for every stock-date in the test window. Look up any ticker to see where its predictions landed. Pooled scatter plots understate a cross-sectional signal — the per-date statistics above are the right lens — but nothing is hidden.
Predicted and realised returns over the forecast horizon, in percent; the realised return is the raw market return of the stock. The regression line is fitted on the full data set, not on the plotted sample.
The full artifact set is available to qualified allocators
Per-date position files, the capacity analysis with square-root market impact at each AUM level, the cost-model calibration against live executions, and the daily signal history through the API are available under NDA. The live signals themselves are published every trading day on this platform, so the track record accrues in public from here on.
Simulated results. The statistics on this page are out-of-sample model outputs applied to a systematic portfolio rule with realised historical costs; they are not the returns of any account and no capital was invested in the simulation. Past performance, simulated or actual, does not guarantee future results. Long/short equity strategies involve leverage, short selling and the potential loss of principal. Nothing on this page is investment advice or an offer of any security or advisory service. Qualified investors should rely on their own due diligence.