Is the combined model always the most accurate?

No, and we do not claim it is. Different models are better at different things: some land closer to the real score, some are more honest about their own uncertainty, some call the direction of the next move more often. We publish the whole comparison rather than crowning one model and hiding the rest.

How many forecast models can I compare on this page?

Eight. Four expect extremes to fade back to normal, differing in whether they also follow recent momentum, adjust what counts as 'normal' for calm versus stressed markets, or take the wider economy into account. One is our own quantum model, which weighs every path the index could take rather than projecting a single line. Three are machine-learning models that read the index's underlying components. All eight are scored on the same test.

Fear & Greed Index forecast — eight models, compared openly

CNN's Fear & Greed Index scores the mood of the market from 0 (extreme fear) to 100 (extreme greed). We forecast where it goes next with eight different models, and we show you how each one has actually performed rather than picking a favourite and hiding the rest.

Four of them expect extremes to fade: they assume an unusually fearful or greedy reading tends to drift back toward normal, differing in whether they also follow recent momentum, adjust what counts as 'normal' for calm versus stressed markets, or take the wider economy into account. One is our own quantum path-integral model, which scores every path the index could take from here instead of projecting a single line. The remaining three are machine-learning models that learn from history how the index's own ingredients — the VIX, market breadth, momentum, the put/call ratio — tend to push the score up or down. All eight are scored on the same test below.

Current Fear & Greed score: 57 (greed). Best-scoring model when tested on data it had never seen: the machine-learning drift model conditioned on the economy (gradient boosting). It missed the real score by about 18.33 points on average, its 90%-confidence range actually contained the real score 82.9% of the time (90% would be perfectly calibrated), and it called the direction of the next move correctly 63.6% of the time.

How the models scored against reality

Every model we offer, tested the honest way: each prediction was made using only the data available at the time, then compared with what the index actually did. They are ranked on two things — how close their stated confidence came to being truthful (a 90%-confidence range should be right about 90% of the time, no more and no less), and how far their forecast landed from the real score. Lower miss is better; confidence closest to 90% is better.

machine-learning drift model conditioned on the economy (gradient boosting) — A machine-learning model that reads wider economic conditions — not just the index's own components — to project which way it drifts. Missed the real score by about 18.33 points on average. Its 90%-confidence range contained the real score 82.9% of the time. It called the direction of the next move correctly 63.6% of the time. Measured over 990 predictions.
machine-learning drift model (random forest) — Learns from history how the underlying components tend to push the reading up or down, and projects that drift forward. Missed the real score by about 17.42 points on average. Its 90%-confidence range contained the real score 78.6% of the time. It called the direction of the next move correctly 72.7% of the time. Measured over 990 predictions.
mean reversion conditioned on the economy — Pulls back toward normal, but what counts as normal shifts with wider economic conditions. Missed the real score by about 18.95 points on average. Its 90%-confidence range contained the real score 74.0% of the time. It called the direction of the next move correctly 75.8% of the time. Measured over 990 predictions.
mean reversion (Ornstein-Uhlenbeck) — Assumes an unusually high or low reading tends to drift back toward its normal level over time. Missed the real score by about 19.15 points on average. Its 90%-confidence range contained the real score 73.9% of the time. It called the direction of the next move correctly 75.8% of the time. Measured over 990 predictions.
mean reversion with momentum (second-order autoregression) — Same pull back toward normal, but it also lets the last few days' direction carry into the near-term path. Missed the real score by about 19.39 points on average. Its 90%-confidence range contained the real score 73.6% of the time. It called the direction of the next move correctly 75.8% of the time. Measured over 990 predictions.
quantum path-integral model with a machine-learning drift — The physics-based path model, with its drift supplied by the machine-learning model rather than assumed. Missed the real score by about 18.23 points on average. Its 90%-confidence range contained the real score 67.7% of the time. It called the direction of the next move correctly 63.6% of the time. Measured over 990 predictions.
quantum path-integral model (Schrödinger) — Our own physics-based model: it scores every path the value could take from here and weights them, rather than projecting one line. Missed the real score by about 20.33 points on average. Its 90%-confidence range contained the real score 66.1% of the time. It called the direction of the next move correctly 66.7% of the time. Measured over 990 predictions.
regime-switching mean reversion — Fits a separate 'normal' level for calm, ordinary and stressed markets, on the view that what counts as normal changes with conditions. Missed the real score by about 18.91 points on average. Its 90%-confidence range contained the real score 61.4% of the time. It called the direction of the next move correctly 69.7% of the time. Measured over 990 predictions.

Frequently asked questions

How is the Fear & Greed Index forecast built?: Two ideas, combined. The first expects extremes to fade: an unusually fearful or greedy reading tends to drift back toward normal. The second is a machine-learning model that learns from history how the index's own ingredients — the VIX, market breadth, momentum, the put/call ratio — tend to push the score up or down. Every model on the page is scored by predicting only from data available at the time and comparing with what actually happened. Educational research; not investment advice.
Is the combined model always the most accurate?: No, and we do not claim it is. Different models are better at different things: some land closer to the real score, some are more honest about their own uncertainty, some call the direction of the next move more often. We publish the whole comparison rather than crowning one model and hiding the rest.
How many forecast models can I compare on this page?: Eight. Four expect extremes to fade back to normal, differing in whether they also follow recent momentum, adjust what counts as 'normal' for calm versus stressed markets, or take the wider economy into account. One is our own quantum model, which weighs every path the index could take rather than projecting a single line. Three are machine-learning models that read the index's underlying components. All eight are scored on the same test.

Educational research only — not investment advice.