2610gnam05/polypulse2-cross-platform-demand

PolyPulse 2.0

Cross-Platform Prediction Market Demand Intelligence

PolyPulse 2.0 is an AI-powered market intelligence system designed to evaluate potential trading demand across prediction-market platforms.

Built on a fine-tuned DeBERTa-v3-base architecture, the model analyzes proposed market questions and classifies their expected relative trading demand into three categories: Low, Medium, and High.

It supports both Polymarket and Kalshi using a single shared neural network with platform-specific decision adjustments.

The model powers the Before Launch — Market Opportunity Intelligence capability within the broader PolyPulse platform.

Model Overview

Attribute	Description
Model architecture	DeBERTa-v3-base
Base model	microsoft/deberta-v3-base
Model type	Transformer-based sequence classification
Supported platforms	Polymarket, Kalshi
Input	Market question, platform and optional event context
Output	Three-class relative trading-demand prediction
Classification labels	Low, Medium, High
Training approach	Supervised fine-tuning with cross-entropy loss
Decision strategy	Validation-selected platform-specific logit adjustments
Primary use case	Pre-launch market-demand assessment

Business Applications

PolyPulse supports market intelligence and opportunity assessment by enabling users to:

• Evaluate market potential: Estimate whether a proposed prediction-market question is likely to attract relatively low, medium or high trading demand.
• Prioritize market opportunities: Compare candidate market questions and identify those with stronger predicted demand.
• Analyze multiple exchanges: Apply the same predictive architecture to Polymarket and Kalshi.
• Support data-driven decisions: Combine demand predictions with historical market comparisons in the PolyPulse application.

The model is designed as a decision-support tool rather than an autonomous trading system.

Model Architecture

The classifier uses a shared DeBERTa-v3-base encoder with a three-class classification head.

Inference workflow:

1. Receive the selected platform and proposed market question.
2. Construct a platform-aware text input.
3. Tokenize and process the input using the fine-tuned DeBERTa model.
4. Generate classification logits for Low, Medium and High demand.
5. Apply validation-selected platform-specific logit adjustments.
6. Return the predicted demand classification and associated model scores.

The same fine-tuned model weights serve both platforms. Platform-specific decision settings account for differences in the historical demand distributions.

Input Format

The expected input follows this structure:

Platform: kalshi | Question: Will the Federal Reserve cut interest rates next month? | Event: Federal Reserve

The model supports English-language market questions with a maximum tokenized sequence length of 128.

Output

The model produces one of three labels:

• Low: Trading demand in the lower range of the platform's historical training distribution.
• Medium: Trading demand in the intermediate range.
• High: Trading demand in the upper range.

The classification thresholds were established using training data separately for each platform.

Model scores are not calibrated probabilities of financial success, future market resolution or investment returns.

Training Data and Methodology

The model was fine-tuned using historical prediction-market records from Polymarket and Kalshi.

The training data includes market questions, platform identification, event context where available, and observed market trading volume.

Because trading-volume units differ between platforms—USD-denominated reported volume for Polymarket and contracts for Kalshi—demand labels are assigned relative to each platform's training distribution.

Training-data quartile thresholds distinguish Low, Medium and High relative demand.

Data Preparation

The preparation workflow incorporates:

• Market eligibility and data-quality checks.
• Historical market filtering.
• Market and event deduplication.
• Chronological training, validation and test partitions.
• Event- and question-level leakage controls.
• Platform-specific volume normalization through classification thresholds.

Fine-Tuning

Two training configurations were evaluated:

1. Standard cross-entropy fine-tuning.
2. Mild class-weighted fine-tuning.

The selected configuration was standard cross-entropy, trained for three epochs with a learning rate of 2e-5, effective batch size of 16, and maximum sequence length of 128 tokens.

Checkpoint and decision-strategy selection were based on validation performance, with accuracy as the primary selection metric.

Model Performance

The evaluated deployment pipeline combines the fine-tuned DeBERTa model with validation-selected platform-specific decision adjustments.

Classification Performance

Evaluation metric	Polymarket	Kalshi	Combined
Accuracy	66.93%	80.43%	70.47%
Macro-F1	0.550	0.706	0.642
Test samples	3,000	1,063	4,063

Benchmark Comparison

Approach	Combined test accuracy
Majority-class baseline	26.36%
TF-IDF + Logistic Regression	54.22%
Fine-tuned DeBERTa with decision adjustments	70.47%

The complete DeBERTa prediction pipeline outperformed the evaluated TF-IDF baseline on both platforms.

Evaluation qualification: The reported 70.47% accuracy includes platform-specific logit adjustments, not just the raw neural-network predictions. The raw model alone achieved approximately 43.76% accuracy on the same test set.

The historical test partition had also been inspected in an earlier experiment. These results should therefore be considered provisional until confirmed on a fresh, independently collected prospective holdout.

Platform-Specific Decision Configuration

The validated decision offsets are applied to the model's raw logits before the final classification.

Platform	Low offset	Medium offset	High offset
Polymarket	+1.0	0.0	-0.5
Kalshi	+1.0	0.0	0.0

The complete deployment configuration is recorded in validation_selection_frozen.json.

To reproduce the evaluated results, these adjustments must be applied consistently during inference.

Limitations

As with any historical-data-driven predictive model, performance depends on the representativeness of the training data.

Known limitations include:

• Predictions are based primarily on market text and platform context, rather than comprehensive real-time market conditions.
• Historical trading patterns can change as platforms evolve.
• The model is less effective at identifying High-demand Polymarket markets than High-demand Kalshi markets.
• Market visibility, external news, liquidity, timing and platform-specific incentives can affect demand independently of question wording.
• Model scores are not independently calibrated outcome probabilities.
• The current evaluation does not establish performance across all future market categories or market conditions.

The model is not designed to predict YES/NO settlement outcomes, determine fair contract prices, or provide investment advice.

Integration within PolyPulse 2.0

The classifier is one component of the broader PolyPulse market-intelligence platform.

PolyPulse also incorporates:

• MiniLM semantic retrieval: Identifies comparable historical prediction markets.
• TimesFM 2.5 forecasting: Provides experimental time-series forecasts of post-launch trading activity.
• Cross-platform analysis: Enables users to evaluate Polymarket and Kalshi separately or through comparative views.
• Interactive market intelligence: Presents predictions, comparable markets and historical trading activity through a Streamlit dashboard.

MiniLM and TimesFM are separate pretrained components. They are not part of this DeBERTa model's uploaded weights or its classification accuracy figures.

Model Repository

Model ID: 2610gnam05/polypulse2-cross-platform-demand

Framework: Hugging Face Transformers / PyTorch

Product: PolyPulse 2.0 — Cross-Platform Prediction Market Intelligence

Model status: Integrated into the PolyPulse application; ongoing validation and monitoring recommended.

---

PolyPulse 2.0 — Turning prediction-market data into actionable market intelligence.