How Stradion works.

For a public set of large-caps, per symbol and horizon: the direction, a calibrated probability of an up move (P(up)), the expected return, and a 90% range. P(up) is calibrated — across the forecasts printed at 62%, about 62% should rise — and 90% interval coverage is published so you can check that against realized outcomes.

The 90% band is the model's uncertainty, not a target. A wider band means less conviction. Stradion publishes only these model outputs and how they scored — no prices, OHLCV, or redistributed vendor data.

Each call is graded after its horizon closes, against only the data the model had at the time — a walk-forward, out-of-sample replay. Nothing is dropped: directional hit rate and 90% coverage are published per model and per horizon, misses included.

Our own trading analysis finds that at short horizons the models' directional confidence does not rank outcomes well enough to trade on; it begins to separate around the 20-day horizon. That result is stated on the track record, not buried. Small samples are noisy — read early numbers alongside the count.

A matrix is a fixed set of symbols, the models that forecast them, and the live record of how those forecasts scored. The public matrix is ten large-caps — eight US names plus Ericsson and Volvo — kept small and free so the record stays legible while it builds. A separate matrix covers the Stockholm OMXS30, and further matrices, each with its own scored record, are planned.

Two forecast models run side by side; each symbol is served by whichever scored better for it out of sample (direction hit-rate, with band calibration as a guard). deep-xgb-cqr is an XGBoost quantile model with conformal bands and a calibrated direction; vol-baseline-v1 is a parsimonious volatility model with no learned features. The track record breaks down each.

The Arena runs the strategies and models as competing paper books, each scored against its market benchmark — SPY for US books, an equal-weight OMXS30 basket for Swedish ones — and ranked by excess return. Three kinds of entrant compete: the strategy champions, each forecast model traded mechanically as a long/short book (long the highest P(up), short the lowest, per horizon), and an LLM agent that picks a book from the published signals and news alone.

It splits into two divisions that are never ranked against each other. Live books are forward-paper, marked daily in real time. Backtest books are walk-forward out-of-sample replays — honest, but never marked live; they graduate to Live only as real forward marks accrue. Every book carries a Deflated Sharpe (the Sharpe corrected for how many configurations were searched, the sample length, and non-normality), and the Arena reports one probability of backtest overfitting (PBO, via combinatorial cross-validation) over the model grid. A book is promoted through explicit stages — candidate, paper-shadow, eligible — only by clearing positive excess, a significant Deflated Sharpe, a PBO under the ceiling, and a minimum live track.

Everything is paper, returns only, no price levels. The Deflated Sharpe's trial count is a documented estimate and its Sharpe-spread is inferred from the visible books, so treat it as indicative, not exact — it is labelled as such in the Arena.

A daily plain-language brief for each watch list — large-cap, income and defensive, megacap tech, and semiconductors, which overlap. A symbol in more than one list is summarized once and shared, so the briefs stay consistent. They are generated from public headlines and can inherit source errors or overstate routine coverage; treat them as a research index, not a trading instruction. Free and anonymous.

Everything published is retained append-only: every forecast, the feature snapshot that produced it, and the score it received once its horizon matured. Records are not revised after the fact. The track record exports in full as CSV — every call, predicted vs realized, including the misses — no account, no email.

The signal matrix, the daily briefs, the Arena, and the full out-of-sample track record are open and anonymous — free examples of what the platform does. The full research desk lives in the Oikos terminal, by invitation; any future member features will live there, not as a tier bolted onto these public surfaces.

Stradion is research, not advice. It does not know your goals, capital, or constraints, and nothing here is a recommendation to buy, sell, or hold. Models degrade when regimes shift, backtests can flatter, small samples mislead, and market data can be delayed or incomplete. Probabilistic forecasts are wrong a stated fraction of the time — a model right 57% of the time is wrong the other 43%. Use it to form your own view.

Stradion began as a private tool: the terminal its author wanted for his own market work, held to the standard he would have paid for. It was built for personal use, not for a buyer, which is why the method and the limits are documented in full.