Home › Insights › How the AI Forecast Works

Methodology

How the Tiresias AI Forecast Predicts Magic: The Gathering Card Prices

Published June 22, 2026 · By Project Tiresias

Project Tiresias publishes an AI price forecast for more than 57,000 Magic: The Gathering singles. Here is exactly how the model turns raw market history into 7, 14, and 30-day projections — and what those confidence ratings actually mean.

Anyone can tell you a card already spiked. The hard part of MTG finance is seeing the move before it happens — and doing it consistently across tens of thousands of cards, not just the handful everyone is already talking about. That is the problem Project Tiresias is built to solve. This page walks through the pipeline behind every forecast on the site.

Why forecasting Magic prices is hard

Magic singles are a thin, volatile, event-driven market. A single reprint announcement can halve a card overnight; a Pro Tour result can double one in a weekend. Most cards trade infrequently, so prices are noisy. And demand comes from several directions at once — competitive play, Commander, collectors, and speculation — that don't always move together. A useful model has to handle all of that without overreacting to noise.

Step 1: Daily price history for the whole market

We log daily prices for every English-language paper Magic single worth tracking — more than 57,000 cards across 700+ sets. Because the model owns its own time series rather than a single snapshot, it can measure the things that actually predict price moves: momentum, volatility, liquidity, and mean-reversion. A card drifting up on steady demand looks very different from one that spiked on a one-day buyout, and the history is what tells them apart.

Step 2: A machine-learning quantile model

At the core is a gradient-boosted (LightGBM) quantile model. Instead of guessing a single future price, it projects a range for each card and horizon — the 10th, 50th, and 90th percentile outcomes. The median (50th) is the headline forecast; the spread between the low and high is where the confidence rating comes from. A tight range means the model is confident; a wide one means the future is genuinely uncertain and you should weight the call accordingly.

In back-testing, the model beats a naïve "tomorrow looks like today" baseline at the 7 and 14-day horizons — which is the honest bar any forecaster should clear before asking you to trust it.

Step 3: A catalyst overlay

Prices don't move in a vacuum, so the statistical forecast is adjusted by an event overlay for the things history alone can't see:

• Reprints & supply. A card spoiled for an upcoming set or Commander deck gets marked down — new supply almost always pushes prices lower. (More on this in our MTG reprint risk guide.)
• Metagame results. Tournament and Commander play data flags cards rising in demand before the broader market catches on.
• Social demand. Spikes in attention and buzz feed into the short-horizon forecasts, where hype matters most.

How to read a Tiresias forecast

The three horizons

Every card shows a 7, 14, and 30-day projection. A card can be flat at 7 days and up sharply at 30 — match the horizon to how long you're actually willing to hold.

The confidence rating

High, medium, or low — derived from how tight the model's predicted range is. A high-confidence +15% is a very different bet from a low-confidence one. Weight your conviction to the model's.

The 80% range

Each forecast shows an 80% interval (the 10th–90th percentile band). The single number is the most likely outcome; the band is the realistic spread around it. Treat the range, not the point, as the real forecast.

What the model can't do

No forecast can price in a surprise — an unannounced reprint, a sudden ban, a rules change, or a viral moment. The model is probabilistic by design: it's right more often than a coin flip, not right every time. It is a tool for making better-informed decisions, not a guarantee, and it is explicitly not financial advice.

Frequently Asked Questions