Sixcents — BTC 15-Minute Model

Quarter-hour predictions are locked to :00, :15, :30 and :45. Each prediction asks one question: will BTC/USD finish the 15-minute window above or below the price locked at the window start?

Predicted 15-minute close

—

Direction & calibrated confidence

CALIBRATING…

Loading historical data and fitting the probability model.

Locked start price

—

Live BTC/USD

—

Window

—

Time remaining

—

Automated historical calibration

Preparing 12 months of 15-minute BTC/USD candles…

Historical windows

—

70% train / 30% test

—

Model accuracy (test)

—

Baseline accuracy (test)

—

Log-loss skill vs baseline

—

Brier: model / baseline

—

The split is chronological, not randomized. The model is fitted only on the first 70% and then frozen before the last 30% is evaluated.

Walk-forward test

A stricter check than the single split. Starting at the 70% mark, the model is refit on everything before each block, then predicts only the next block it has never seen. The window rolls forward until the data ends, and all blocks are scored together.

Refits / windows scored

—

Model accuracy

—

Baseline accuracy

—

Log-loss skill vs baseline

—

Brier: model / baseline

—

Folds beating baseline

—

Learned feature weights

Each feature is standardized on the training set, then logistic regression learns its weight. A positive weight means higher values push toward "Higher"; a negative weight means the feature has tended to mean-revert.

Weights appear after calibration.

No-look-ahead design

For historical window i, the locked price is that 15-minute candle's open and every feature is calculated only from candles that ended before window i began. The candle's close is used only afterward as the Higher/Lower outcome. Feature scaling, weights and the price-move table all come from the training 70%; the test 30% never participates in fitting.

Recent locked windows on this device

No completed local predictions yet.