Prediction Markets & Corporate Earnings:
An Accuracy Analysis

Key Findings

• 79.7% overall directional accuracy at D-1 (the day before earnings) across all 909 events
• When markets priced a >80% or <20% probability (high-conviction signals), accuracy rose to 87.8%
• 63% of events had already reached high-conviction pricing (>80% or <20%) by D-1
• The naive baseline (always predicting "Beat") yields only 72.6% — markets add meaningful lift beyond this floor
• Miss detection is significantly harder: markets correctly flagged only 45.2% of eventual misses at D-1, while identifying 92.7% of eventual beats
• Higher-volume markets showed considerably better accuracy (87.6%) vs. lower-volume markets (71.8%)
• Total market volume across 909 events: $11.3 million

1.1 What the markets measure

Each Polymarket earnings market poses a simple binary question: "Will [Company] beat quarterly earnings?" Contracts trade between $0 and $1. A price of $0.72 implies a 72% market-assigned probability of an earnings beat. Resolution occurs at the closing price: contracts near $1.00 (≥$0.98) resolve as "Beat," while contracts near $0.00 (≤$0.02) resolve as "Miss."

1.2 Structural differences from Fed rate markets

Fed rate decision markets typically have 4–8 weeks of active trading before resolution. Earnings markets are structurally different: most companies report quarterly, and prediction market trading tends to concentrate in the final 7–14 days before the earnings announcement. The median event in this dataset had only 11 days of pricing data, and 25% had 8 or fewer data points.

This means the signal-building process documented in the Fed analysis (confidence rising steadily from 30+ days out) is largely absent for earnings. The market dynamic is more abrupt: traders enter a concentrated window, form views, and the market resolves quickly.

1.3 Dataset overview

This analysis covers 909 completed earnings beat/miss markets traded on Polymarket from June 2025 through May 2026. The dataset captures a broad cross-section of U.S. publicly traded companies — from mega-cap technology names (NVIDIA, Apple, Alphabet, Meta, Microsoft) to mid-cap industrials, consumer companies, and speculative growth stocks. Total traded dollar volume across all markets was $11.3 million.

2.1 The base rate problem

Any assessment of prediction market accuracy for earnings must account for the "base rate" effect. Over the Jun 2025 – May 2026 period, 72.6% of companies beat estimates — a figure consistent with the historical tendency for companies to manage guidance conservatively. A trivial "always predict Beat" model would therefore achieve 72.6% accuracy with zero effort.

The meaningful question is whether prediction markets deliver lift over this baseline — and whether they correctly identify the minority of companies that will miss. The answer is nuanced: the markets do add lift, but unevenly.

2.2 Accuracy by time horizon

The table below shows directional accuracy and average implied confidence at each time horizon before earnings. Unlike Fed rate markets — where confidence built steadily over 30+ days — earnings markets show relatively flat accuracy from D-7 to D-14, with a meaningful jump at D-1 as final pricing converges.

High-conf. accuracy = accuracy for events priced >80% or <20% at that horizon. D-30/D-21 sample sizes are small due to most markets not opening that early.

2.3 The high-confidence signal

The most actionable finding in this data is the divergence between all events and high-confidence events. When the market priced a beat probability above 80% or below 20%, accuracy at D-1 jumped to 87.8%. Roughly 63% of all events fell in this category by D-1 — meaning a majority of markets were giving clean, high-confidence signals.

3.1 The beat/miss asymmetry

The most important structural feature of earnings prediction markets is the skew: companies beat estimates far more often than they miss. Over this dataset, 72.6% of companies beat. This base rate creates a systematic bias — if uncertain about an outcome, the rational prior is "probably a Beat." The markets reflect this: the average D-1 beat probability across all events was approximately 70%.

The performance profile is dramatically different for beats vs. misses. Markets were highly effective at identifying eventual Beats (92.7% recall) but caught only 45.2% of Misses at D-1 using the 50% threshold. In other words, when a company ultimately missed, the market had it above 50% — predicting a beat — nearly 55% of the time.

3.2 The sub-60% signal: precision vs. recall tradeoff

Rather than using the 50% threshold mechanically, a more actionable framing asks: when the market prices a company below 60% to beat, how informative is that? The answer depends on how far below 60% — and which sector the company is in.

The sub-20% bucket is the cleanest signal: 88.5% precision with only 6 false alarms in 52 flagged events. But it catches fewer than 1-in-5 actual misses. The uncomfortable reality is that 46.4% of all misses were priced above 60% at D-1 — the market simply did not anticipate them. These are genuine surprise misses.

The price trajectory also matters: for events that ultimately missed, the average D-7 beat probability was 0.60, falling to 0.51 by D-1. About 58% of eventual misses saw their price fall between D-7 and D-1 — a declining trend that itself is a secondary signal, independent of the absolute level.

3.3 Miss rates by sector: where the signal is strongest

Miss rates vary dramatically across sectors, and so does the quality of the sub-60% warning signal. The data reveals three distinct tiers:

Surprise Misses = % of actual misses that were priced above 60% at D-1 (the market did not anticipate them). Sub-60% Precision = of events priced below 60%, % that actually missed. Sub-60% Recall = % of all misses that were flagged by the sub-60% signal.

Three patterns stand out:

Real Estate: the most reliable miss signal in the dataset. With a 71.4% base miss rate and 90% sub-60% precision and recall, Real Estate markets appear to price in distress quite accurately — likely because deteriorating fundamentals in property-linked companies tend to be visible in advance. Only 10% of misses came as surprises.

Communication Services and Healthcare: strong signal, moderate miss rates. Both sectors showed sub-60% precision above 70% and recall above 70%, meaning when the market flagged these names as at-risk, it was usually right and it caught most of the misses. These are the sectors where the sub-60% screen is most actionable.

Technology: low miss rates but mostly surprise misses. Technology companies beat at a remarkable 90.2% rate — the highest in the dataset. When they did miss, however, 56% of those misses were priced above 60% at D-1, meaning the market was caught off guard. Tech is a sector where the prediction market is primarily useful as a beat confirmer, not a miss detector.

Consumer Staples and Industrials: the market adds least value. Both sectors had low base miss rates (around 20%) and poor sub-60% signal quality — with recall below 35% and high surprise miss rates (69–79%). These sectors may resist prediction because their earnings surprises stem from idiosyncratic factors (input cost shocks, weather, supply chain) rather than the kind of public information that prediction market participants can synthesize.

3.4 Miss rates by market cap: smaller companies are harder to predict

Market capitalization shows a clean, monotonic relationship with both miss rates and the quality of the miss detection signal. Smaller companies miss more often, and when they do miss, the sub-60% signal is sharper.

Small-cap companies missed at nearly three times the rate of mega-caps (45.7% vs. 14.1%). More importantly, the sub-60% signal for small-cap names had 72.7% precision and 64.9% recall — the best of any tier — meaning the market was better at anticipating small-cap misses than large-cap ones. This is plausible: small-cap companies tend to have higher earnings volatility and weaker guidance frameworks, making genuine uncertainty more visible in market pricing.

Large-cap stocks showed the weakest miss signal precision (55.7%) with the highest surprise miss rate (54.7%). For mega-cap and large-cap names, the market tends to assume a beat until proven otherwise, and misses often arrive as genuine shocks — consistent with the idea that large companies have more sophisticated investor relations and better ability to manage analyst expectations.

3.5 Companies that missed repeatedly

Several companies appeared multiple times in the dataset and missed earnings on every event tracked. While sample sizes are small, persistent missers may represent structurally unpredictable businesses or companies in operational distress during this period.

4.1 Volume as a quality signal

Dollar volume varies enormously across the 909 markets — from under $100 for the thinnest names to over $650,000 for the most-watched events (NVIDIA). This variation is itself informative: high-volume markets attract more sophisticated participants and are likely to be better calibrated.

Splitting the dataset at the median volume, higher-volume markets achieved 87.6% D-1 accuracy vs. only 71.8% for lower-volume markets — a 15.8 percentage point gap. This strongly suggests that volume is a useful proxy for signal quality: thinner markets should be weighted less in any investment framework.

4.2 Highest-volume markets

The table below shows the 15 most actively traded earnings markets in the dataset. These are the events where prediction market data is likely most reliable. NVIDIA's earnings events alone accounted for nearly $900,000 in combined volume.

4.3 Notable prediction errors

The most notable high-volume mispredictions at D-1 reveal a pattern: they cluster around events with genuine uncertainty — where both beats and misses were plausible given the public information at the time. The market was not blindly wrong; in many cases it was expressing real uncertainty that happened to resolve the other way.

Tesla appears multiple times in this table — including once with 83% confidence in a Beat that turned into a Miss, and once with 71% confidence in a Miss that turned out to be a Beat. Tesla earnings have long been idiosyncratic and difficult to forecast, and this data confirms the prediction market also struggled with them.

5.1 Beat rate variation by earnings season

The underlying beat rate varies across reporting periods. The dataset spans three major earnings seasons (Q3 2025, Q4 2025, and Q1 2026), each with a slightly different beat rate. Markets must implicitly calibrate to these seasonal patterns.

The Q1 2026 reporting season showed a notably lower beat rate (68.8%) compared to Q4 2025 (75.1%), reflecting broader macro uncertainty in early 2026. Prediction markets tracking these events in real time would have needed to implicitly adjust for changing base rates — a sophisticated task that underlines why even modest lift over the naive baseline is meaningful.

5.2 Calibration quality

The cleanest test of calibration is this: do events priced at 80%+ actually beat at an ~80%+ rate? In this dataset, events at D-1 split almost entirely into two buckets — <10% (all misses) and >90% (all beats) — with very few intermediate prices at resolution. This binary polarization is expected: markets settle into near-certainty by the time of resolution. The more interesting calibration signal is in the intermediate prices 7–14 days out, where the market is still forming its view.

6.1 Strategies informed by earnings prediction market data

6.2 Comparison to Fed rate market performance

The contrast with Fed rate decision markets is instructive. Fed markets achieved 95.7% directional accuracy on decision day — but that was in a structured, calendar-driven environment with extensive Fed communication, FOMC transcripts, and blackout periods that give markets a clear cadence to work with.

Earnings markets operate in a noisier environment: management teams control the information flow, guidance is strategic, and macro surprises can invalidate even well-researched forecasts. An 87.8% accuracy rate for high-confidence earnings signals is therefore arguably as impressive as 95.7% for Fed decisions — the underlying task is harder.

6.3 Limitations and caveats

Several important caveats apply to this analysis:

• Definition of "Beat": Resolution here uses the final prediction market price, which itself reflects consensus on whether the company beat or missed. This is self-consistent but may not always align with every analyst's definition.
• Limited pricing history: Most events have only 7–14 days of data. With more lead time, signal quality might improve — or the market might remain uncertain until very close to the announcement.
• Adverse selection: The fact that a prediction market exists for an event may itself be endogenous — markets are more likely to be created for earnings events perceived as interesting or uncertain.
• Short time series: This analysis covers approximately one year (Jun 2025 – May 2026). A longer dataset would allow for more robust calibration analysis and seasonal pattern identification.