Meet the Oracle
The Oracle is the machine-learning model behind every trade you take. It never pulls the trigger for you. It hands you one honest read on each ticker you watch, a few times a day, so you can stop second-guessing and act with an edge. Here is exactly what goes into that read.
Every card carries seven decisions
Scroll slowly. The card stays put while it comes apart, one piece at a time.
- 1Calls or puts
The Oracle is symmetric. It reads downside as readily as upside, so a bearish setup surfaces as BUY PUTS, not just a shrug.
- 2Raw ensemble confidence
Five independent engines - trend, institutional flow, momentum, news sentiment, and a prediction layer - vote on every ticker. Their weighted blend is this raw score.
- 3Calibrated win rate
The raw score is re-mapped against thousands of real outcomes, so this number tracks how often signals at this confidence actually worked out, not how loud the model is.
- 4Conviction tier
A, B, or C grades the setup. The tier scales position size with a half-Kelly rule, so the model leans in only when the edge is clearer.
- 5Live and time-stamped
Every card carries the real-time price and the exact second it was computed. Signals refresh through the session, not once a day.
- 6Regime aware
It reads the wider market first. In a bull tape it leans in; in a bear tape it asks for more evidence before it says buy.
- 7The next five minutes
A short-horizon read layered on top: a quick lean for the minutes just ahead, learned from how similar setups resolved.
Seventeen tickers,
one deck
You cannot stare at the whole market at once. The Oracle can. It reads every ticker you follow and deals you a fresh card for each. Scroll to thumb through the stack.
Seventeen tickers, one deck
Five engines, one vote
There is no single black box. Five independent engines each score the same ticker from a different angle, and the Oracle blends them by weight. No engine can carry a trade alone.
The number you can trust is the calibrated one
A raw model score is easy to inflate. So the Oracle re-maps every score against thousands of past outcomes with isotonic regression. A calibrated 70 means signals at that level actually resolved in your favour about seven times in ten. That is the small emerald line on the card, and it is the number that actually matters.
How loudly the ensemble is shouting. Useful, but it can be miscalibrated.
The same score, corrected against real history. This is the honest read.
It reads the room, then sizes the bet
Before it scores a single ticker, the Oracle classifies the broader market as bull, bear, or neutral from the index trend and volatility. In a bear tape it raises the bar to call a buy and lowers it to call a sell. The conviction tier then translates directly into size.
Clearest edge. Largest position under a half-Kelly rule.
The everyday signal. A measured, baseline position.
Worth noting, sized small. Often skipped entirely.
A quiet loop that runs every night
After the close, the model re-grades itself on the day it just lived through, retrains a challenger, and tests it on data it has never seen. The challenger only takes the crown if it genuinely beats the reigning model out of sample, after trading costs. Each champion gets a name, so upgrades are easy to track over time.
Score the freshest outcomes and grow the training set.
Train a rival and test it on untouched, future data.
Swap only if the rival wins after costs. Otherwise, no change.
Graded the way a skeptic would
We would rather show you the method than a flattering headline number. Here is how every version of the model earns its place.
Each signal is labeled by whether a fixed profit target, a fixed stop, or a time limit hit first. No cherry-picking the exit.
The model trains on older data and is scored only on newer data it never saw, so results are not memorised.
Commissions and slippage are subtracted before anything counts. The bar is beating the cost of trading, not zero.
Predicted confidence is plotted against what actually happened. The closer to the diagonal, the more honest the number.
A calibration curve is the plainest test of an honest model. Each point compares what the model predicted with what really happened. The dashed line is perfect calibration. The whole aim of the calibration step is to pull the model onto that line. The axes here carry no numbers on purpose, because this shows the method, not a performance claim.
A note on honesty: the Oracle is a research tool, not financial advice, and not a promise. Signals are one input into a human decision. Markets change, models drift, and past behaviour does not guarantee future results.
See the Oracle read a live ticker
Watch the card update in real time and decide for yourself. Free during the open beta.