← Back to forecasts

Methodology

Every evening after the U.S. close, each model in the roster is asked one question per ticker: where will this security close at the end of the next trading session? This page describes how those answers are produced, combined and scored.

Data sources

Models use split- and dividend-adjusted daily bars (open, high, low, close, volume) from Yahoo Finance for every covered security, plus SPY, QQQ, the VIX and the 10-year Treasury yield as cross-asset inputs. A partial bar for a session that is still trading is discarded, so no model sees data from after the close it forecasts from.

The model roster

Each model is one algorithm paired with one feature set. Twenty-five machine-learning algorithms — gradient boosting (including XGBoost and LightGBM), tree ensembles, regularised linear models, kernel methods, a neural network, Bayesian regressions and two stacked ensembles — are each run on three feature sets: price and volume, technical indicators, and cross-asset returns. They predict the next session’s return. Seven classical time-series methods (ARIMA, autoregression, exponential smoothing, damped Holt, Theta, drift and momentum) forecast the next close from the price series alone. That makes 82 models from 32 distinct algorithms per ticker.

How models are trained and scored

Feature models train on the most recent 504 sessions (252 for the Gaussian process). Every model is scored walk-forward over the last 30 sessions: for each session it is fitted only on data available the evening before, then compared with the actual close. Feature models are refitted every five sessions in that test and once more for the published forecast; time-series models are refitted every session. Hyper-parameters are fixed in advance and are not tuned on the scored sessions.

How the consensus is formed

The headline figure is the median of all reporting models, because the median is not dragged by a few extreme forecasts. We also publish the mean, an accuracy-weighted mean (weights proportional to each model’s trailing 30-day directional accuracy above 40%), the interquartile range and the full min–max range. The 30-day record shows the median of the models’ walk-forward forecasts for each past session.

Confidence and agreement

A model’s confidence compares its average error over the last 30 sessions with that of simply predicting no change: 50 means no better, higher means smaller errors. Its interval is an 80% band built from the same errors. Agreement measures how narrow the interquartile range is relative to the last close. The headline confidence blends average model confidence (50%), agreement (30%) and the lopsidedness of the up/down split (20%). None of these is a probability that the forecast is correct.

Update schedule

The pipeline runs once per trading day after the U.S. close and forecasts the following session. Forecasts are not revised overnight. On weekends and exchange holidays the page shows the forecast for the next valid session, and a banner appears if the latest forecast is for a session that has already closed.

Known limitations

Next-day price moves are close to unpredictable. Across the roster, directional accuracy typically sits near 50%, and a model that did well last month often regresses. Models cannot anticipate earnings surprises, news or gaps, and many share inputs, so their errors are correlated — 82 models agreeing is weaker evidence than it looks. The consensus is frequently wrong, and the 30-day record on every ticker page shows the misses alongside the hits.