Can Google’s TabFM Predict Polymarket Crypto Markets? A Zero-Shot Backtest

Executive Summary

Polymarket’s 5-minute crypto contracts ask a deceptively simple question: will price be higher or lower five minutes from now? They settle fast enough that there is nowhere to hide a bad model. This report tests whether Google Research’s TabFM (Tabular Foundation Model) can compete with trained classifiers on that question, using a strict, leakage-free walk-forward framework across 1,320 prediction points on BTC, ETH, and XRP.

The headline result: TabFM’s zero-shot predictions (no training, no gradient updates, no hyperparameter search) hit 53.86% accuracy on XRP, edging out a fully trained XGBoost baseline (51.97%) on that asset and landing within a point of the best model on every symbol tested. A companion analysis of Logistic Regression coefficients, Random Forest importances, and SHAP values across all three assets converges on the same three drivers of these markets: short-term momentum, medium-term trend exhaustion, and intra-session mean reversion around VWAP.

Figure 1: SHAP summary for XRP, the model’s best-performing asset. Full breakdown of what drives these predictions in Section 5.3.

1. The Prediction Challenge & Data Alignment

Polymarket’s 5-minute prediction markets for major cryptocurrencies (BTC, ETH, and XRP) resolve UP if price rises over the interval, and DOWN if it falls or stays flat.

Predicting these markets requires a strict, leakage-free data pipeline. For a market spanning 16:00 to 16:05:

1. Predictions must be made at or before the 16:00 start time.
2. Features must use technical indicators calculated strictly from historical data before 16:00.
3. This setup matches the operational constraint of entering a live trade right as the market opens.

This analysis uses a 2-week historical dataset (early April 2026), yielding approximately 1,445 closed markets per asset. Features are extracted across three timescales:

• Micro-scale (1-min): 5-minute rolling window statistics (mean, std, delta, last) of prices, session VWAP, and short-term RSI.
• Medium-scale (15-min and 1-hour): Standard indicators such as RSI, MACD, ADX, and Bollinger Bands.
• Macro-scale (1-day): Long-term trends such as the daily 200-day Simple Moving Average.

2. Foundation Models for Tabular Data: TabFM vs. Text LLMs

Traditional tabular machine learning models require explicit training by updating weights on a specific dataset. In contrast, TabFM is a Tabular Foundation Model that performs zero-shot classification.

Architecture and Inference Model

• Specialized Architecture: Unlike text-based LLMs (e.g., GPT-4 or Llama) that process character tokens and output natural language, TabFM is designed specifically for numerical and categorical tabular data. It maps columns to vector embeddings and outputs class probabilities directly.
• In-Context Learning (ICL): At inference time, TabFM’s weights remain frozen. Instead of parameter training, it reads a historical training set as "context" and makes predictions on test samples in a single forward pass, leveraging row and column attention mechanisms.
• Local Execution: The model runs entirely locally using PyTorch (google/tabfm-1.0.0-pytorch) without external API calls or external text-generation servers.

3. Walk-Forward Backtesting: Folds & Runtime

Models are evaluated using sequential walk-forward validation, which models a real-world trading setup by expanding the training history over time.

Understanding Walk-Forward Folds

A "fold" represents a single backtesting step. For each fold:

1. The model is presented with a historical training context (120 observations, representing 10 hours of market history).
2. The model predicts the outcomes of the next test block (24 observations, representing a 2-hour window).
3. The training window then expands by 24 observations, and the process repeats for the next fold.
4. Across the 2-week dataset, this yields 55 sequential folds.

TabFM Computational Overhead

TabFM requires 7.5 to 8.5 seconds per fold, totaling approximately 7.5 minutes of compute time per asset (22 minutes total for BTC, ETH, and XRP combined).

Unlike traditional models that evaluate simple tree splits or linear equations in milliseconds, TabFM must pass the entire historical context window (120 rows) and the test queries together through a 24-block causal transformer network. On each step, the transformer dynamically calculates row-to-row and column-to-column attention weights, resulting in higher computational latency.

4. Performance Comparison

The table below outlines the backtesting accuracies and F1-scores across 1,320 prediction points:

TabFM’s zero-shot inference is highly competitive: it outperforms the trained XGBoost baseline on XRP (53.86% vs. 51.97%) and matches the top SVM accuracy on ETH, all without ever seeing a labeled training example specific to this dataset.

5. Economic & Quantitative Interpretation of the Models

Analyzing the inner workings of Logistic Regression and Random Forest models reveals the structural dynamics of prediction market behavior.

5.1 Logistic Regression Coefficients: A Quantitative Rationale

The coefficients from the Logistic Regression models represent the change in log-odds of a market resolving UP for a one standard deviation increase in the corresponding feature.

BTC Logistic Regression Coefficients

Figure 5.1: BTC Logistic Regression Coefficients

ETH Logistic Regression Coefficients

Figure 5.2: ETH Logistic Regression Coefficients

XRP Logistic Regression Coefficients

Figure 5.3: XRP Logistic Regression Coefficients

These charts show a highly consistent structure across all three assets:

• Momentum (tf_15min_rsi_14), Strongly Positive (e.g., +0.550 for XRP): Strong price momentum leading up to the market opening is the single strongest indicator of an UP resolution. Economically, this reflects short-term trend persistence: buying interest on a 15-minute horizon tends to carry over into the subsequent 5-minute window.
• Trend Exhaustion (tf_1hour_macd_hist), Strongly Negative (e.g., -0.432 for XRP): The MACD histogram measures the acceleration of price trends. A highly elevated MACD histogram indicates the trend is moving at maximum velocity, which historically represents late-stage exhaustion. On a 5-minute horizon, these overextended trends are highly prone to reverse or consolidate.
• Mean Reversion (tf_1min_mean_5_vwap_session), Strongly Negative (e.g., -0.367 for XRP): When price trades significantly above the session’s Volume-Weighted Average Price (VWAP) in the final 5 minutes, the probability of an UP resolution falls. This is intra-session mean reversion: liquidity providers and arbitrageurs fade extreme short-term price deviations back toward the average session value.

5.2 Random Forest Feature Importances

The Gini importances of the Random Forest model confirm the importance of these same variables.

BTC Random Forest Feature Importances

Figure 5.4: BTC Random Forest Feature Importances

ETH Random Forest Feature Importances

Figure 5.5: ETH Random Forest Feature Importances

XRP Random Forest Feature Importances

Figure 5.6: XRP Random Forest Feature Importances

Short-term momentum (tf_15min_rsi_14) and session VWAP deviation (tf_1min_mean_5_vwap_session) dominate the importance scores, each accounting for 12% to 15% of the total predictive power. Conversely, long-term trends like the daily 200-day moving average (tf_1day_sma_200) account for less than 1% of the model’s split decisions. Macro structural trends have essentially no statistical influence on 5-minute prediction horizons.

5.3 SHAP Summary (Beeswarm Analysis)

TabFM is a zero-shot transformer, so it does not expose the per-feature attributions that SHAP’s TreeExplainer computes for tree-based models; no TabFM-native SHAP values exist. As a proxy, we computed Shapley Additive exPlanations (SHAP) values for the Random Forest model on XRP, the ticker TabFM performed best on, to illustrate the direction of feature impact on individual predictions.

Figure 5.7: XRP Random Forest SHAP Summary (proxy explainer, shown for TabFM’s best-performing ticker)

On a SHAP summary plot, horizontal position represents a feature’s impact on the model output (positive values increase the probability of an UP prediction), and color represents the feature value (red is high, blue is low):

• tf_15min_rsi_14: Red dots cluster on the positive side of the SHAP axis, blue dots on the negative side. High RSI values consistently push predictions toward UP, a clean momentum effect.
• tf_1min_mean_5_vwap_session: Red dots cluster on the negative side, blue dots on the positive side. High deviations above the session VWAP push predictions toward DOWN, validating the mean-reversion driver.
• tf_1min_delta_5_ema_9: This variable measures trend acceleration over the final minutes. High values (red) associate with positive SHAP values, showing short-term trend acceleration acts as a reinforcing signal.

The alignment across Logistic Regression coefficients, Random Forest importances, and SHAP distributions confirms that both linear and non-linear models capture the same underlying market structure: following medium-term momentum, but fading extreme short-term deviations from the session average.

6. Correlation Analysis & Decoupling

The heatmap visualizes the correlation matrix of the key indicators.

BTC Indicator Correlation Heatmap

Figure 6.1: BTC Technical Indicator Correlation Heatmap

ETH Indicator Correlation Heatmap

Figure 6.2: ETH Technical Indicator Correlation Heatmap

XRP Indicator Correlation Heatmap

Figure 6.3: XRP Technical Indicator Correlation Heatmap

Indicators of similar frequencies show moderate correlation (e.g. r ≈ 0.65 between 15-minute and 1-hour RSI), but short-term 1-minute indicators are decoupled from hourly and daily trends. This structure lets the models combine microstructural mean-reversion signals with medium-term directional momentum without suffering from multicollinearity.

7. Conclusion

This study demonstrates that Google Research’s TabFM zero-shot tabular foundation model is capable of predicting short-horizon prediction markets. By leveraging pre-trained representations and in-context learning with a 120-row historical window, TabFM achieved a competitive 53.86% accuracy on XRP, outperforming the fully trained XGBoost baseline on that asset.

Logistic Regression coefficients and SHAP values show that Polymarket outcomes on these 5-minute crypto contracts are driven by a combination of trend momentum (RSI), trend exhaustion (MACD), and mean reversion (session VWAP). Tabular foundation models carry higher computational latency from the transformer’s forward pass, but their zero-shot capability, which skips training entirely, is a genuinely useful alternative to standing up a full model-training pipeline for short-horizon financial forecasting.

None of the accuracies above clear a large enough margin over a 50% coin flip to be treated as a standalone edge before transaction costs, market impact, and Polymarket’s fee structure are accounted for. This is a research screen into what drives these markets, not a validated trading signal.

Limitations

This is a research screen based on a 2-week sample window and a single historical run, not a validated trading strategy. Accuracies in the low-to-mid 50s are directionally interesting but have not been tested against Polymarket’s fee structure, execution slippage, or a longer out-of-sample period. Nothing in this article is investment advice.