Loading…

—

Predicted path

Scroll to zoom · drag to pan · drag the price axis to stretch it · double-click to reset · » jumps back to today

Describes recent price action — not a forecast on its own

Gets live data. A small local server (started by Start Stock Predictor.bat) fetches up to 10 years of daily prices, today’s 5-minute bars and live quotes from Yahoo Finance, refreshing every 30 seconds.
Fits ten forecasting models in the background, from a plain random walk to GARCH volatility models, a Kalman filter and two machine-learning models.
Turns each into a probability distribution of the future price — not one number, but how likely every price is on every future trading day.
Backtests every model on this stock’s own history: dozens of times in the past, each model is re-fitted using only the data available that day, and its forecast is scored against what actually happened.
Combines the models into an ensemble that gives more weight to the ones that did better in the backtest.

The median (dashed line) is the middle outcome: the models think the price is equally likely to end above or below it.
The shaded bands are the 50%, 80% and 95% ranges. An 80% range is expected to miss 1 time in 5. That is normal and honest, not a failure.
“Chance it ends higher” is the share of the forecast distribution above today’s price. Values near 50% mean the models see no real edge, which is the most common and most truthful answer.
The odds calculator gives both the chance of finishing beyond a price and the chance of touching it at any point before the date (estimated from thousands of simulated paths).
Forecasts are anchored to the latest live price, so the fan moves with the market between model refits.

Walk-forward backtest: Forecasts are made at evenly spaced past dates using only data up to each date (after at least three years of training data when available) and compared with the actual price h days later. No future information leaks into any forecast, and the ensemble’s weights are learned only from outcomes that were already known at each date.
Typical error: Root-mean-square error of the median forecast in log terms, shown as a percentage.
CRPS and skill: The continuous ranked probability score rewards a forecast distribution that is both accurate and appropriately confident. Skill = 1 − CRPS(model) ÷ CRPS(no-change). Positive means better than the no-change benchmark.
p-value: Diebold–Mariano test of whether the model’s CRPS differs from the benchmark’s, with Newey–West correction for overlapping forecast windows. Below 0.10 is flagged as significant, but with many models and horizons some “significant” results will be luck.
Direction right: How often the median pointed the right way (up or down). Compare it with the up-move base rate: a model that always says “up” scores the base rate.
Range hit rate (calibration): How often the actual price landed inside the 80% and 95% ranges. Much less than 80% means the model was overconfident; much more means its ranges were too wide.

Stock prices are close to unpredictable. Public information is priced in quickly, and most models struggle to beat “no change” by a statistically meaningful margin. The backtest tab shows exactly how each model did on this stock.
The models only see past prices and volumes. They know nothing about earnings, news, interest-rate decisions, lawsuits or takeovers, and those are often what moves a price.
Real returns have fat tails and sudden regime changes. A model calibrated in calm years can underestimate crash risk.
Machine-learning models are prone to finding patterns in noise. This app regularises them heavily and reports their out-of-sample results honestly.
Prices are split-adjusted but exclude dividends. Some exchanges are delayed by 15–20 minutes. Holidays are modelled for US exchanges; other markets use a Monday–Friday calendar.
This is an educational and analytical tool, not financial advice. Never invest money you cannot afford to lose based on a model’s forecast.

Self-tests for the maths: tests.html