How Candor works, and why it usually says “No edge”

What the numbers mean

Each card shows the model's probability that the price will be higher at the end of the horizon than now, plus a typical range of moves (10th to 90th percentile, scaled to current volatility). It never says “it will go up”.

A card shows No edge when any of these is true: the model failed to beat simple baselines on unseen data for that horizon, its probability is within a few points of 50%, the data is too old for the horizon, or the market is closed.

How it's tested

• Walk-forward only. The model is always trained on the past and tested on later data it never saw. There are no random splits.
• Embargo. A gap equal to the horizon separates training from testing, so no training label overlaps the test period.
• Untouched holdout. The most recent 20% of history is never used to choose a model or a setting. The scores you see come from there.
• Baselines. Coin flip, “always up” (the drift that buy-and-hold earns), “same as the last move”, and buy-and-hold returns, all scored on the same rows.
• Significance. Overlapping horizons aren't independent, so the p-value counts only one prediction per horizon length.
• Costs. Simulated trades pay half the spread plus slippage on each side.
• Shrinkage. Probabilities are pulled toward 50% by an amount chosen on development data. A model with no signal ends up saying ~50% instead of being confidently wrong.
• Simple models first. Regularised logistic regression and shallow gradient boosting. The one with the better out-of-sample log loss in development is used.

Signals

• Technical: momentum, moving averages, RSI, MACD, VWAP distance, volatility, volume spikes, bar range.
• Order flow (approximate): where bars close within their range, and up-tick vs down-tick volume. True Level 2 depth needs a paid feed (see Data sources). With Alpaca keys, the live bid-ask spread is shown and used for paper fills.
• Market context: SPY, QQQ, the stock's sector ETF, relative strength, time of day, overnight gap. VIX and days-to-earnings are used by the daily model only. Intraday history is too short for them, and the model would memorise the days.
• News: headlines are scored with a finance word list and weighted by recency. They enter the model only when timestamped history is available (Alpaca news), so they can be tested. Optional LLM scores are shown for context only. An LLM scoring old headlines already knows what happened next, so backtesting with it would leak the future.

Contracts

• For short-window yes/no contracts (“Will BTC be below $80,391 at 3:15?”), direction barely matters. What matters is how far the price is from the strike compared with how much it typically moves in the time left.
• Two estimates are blended. A normal model uses recent 1-minute volatility. A history model looks at thousands of real past windows of the same length, each scaled by the volatility at the time, which keeps sudden jumps in. Data delay counts as extra time.
• A contract is only called “underpriced” if it's cheap at every point in the uncertainty range, after fees. Most of the time the honest answer is “no edge”.
• Each check is graded after it closes. The scoreboard compares Candor's odds with the market's prices, and shows what following the calls would have earned.

AI analyst

• When you click “Ask the AI analyst”, Google Gemini first searches the web: recent news, analyst moves, upcoming earnings or Fed meetings, sector and macro trends, investor sentiment. A second Gemini model combines that research with Candor's prices, market context and the tested models' results, and gives a probability and a reason for every horizon.
• Why the AI isn't backtested: a language model has read about the past, including what happened after any date we'd test it on. Testing it on history would be cheating. The only honest test is live: every AI call is logged and graded in Track record, side by side with the statistical model and a coin flip.
• Until it has a solid live record, its views are labelled “unproven”. The cards' “No edge” verdicts still come from the tested statistical models.

Crypto

• Search a coin (e.g. “bitcoin”) or type its pair, like BTC-USD or ETH-USD. Coins trade 24/7, so intraday predictions run at any hour, weekends included. Crypto “days” run midnight to midnight UTC, and the daily card becomes “Next 24 hours”.
• Coins other than Bitcoin also use Bitcoin's own recent moves as a signal. SPY/QQQ only change while US markets are open.
• Simulated crypto trades pay a wider spread plus an exchange fee on each side (default 0.25%). Apps with “simple buy” buttons often charge much more.
• Crypto prices here are Yahoo's average across exchanges. Very short-term patterns in averaged prices can be artifacts that don't exist on any real exchange. That's one reason a model can beat the baselines on accuracy and still lose money after costs, and Candor flags exactly that case.

Known limits

• If you test many tickers and horizons, about 1 in 20 will look “significant” by pure luck. Treat a single green check with suspicion, and watch the live track record.
• Yahoo data is unofficial and may be delayed. Market holidays are hard-coded and half-days are not handled.
• Past patterns, even real ones, fade once enough people trade on them.