Oblivia - Price Predictions

Oblivia (our price prediction model) is a very very very experimental feature we rolled out, inspired by users’ requests to take advantage of our on-chain and social data to generate unified price predictions for Solana memecoins.

Today, we have 89% directional accuracy on Solana memecoin price predictions.

Oblivia 1.0 helps forecast price movements by analyzing a blend of social and on-chain. It powers key decision-making across the platform by offering high-confidence directional predictions.

The model achieves 89% direction accuracy across all predictions, with performance improving to 93.6% for high-confidence signals. Notably, price increase predictions are 91.2% accurate.

Predictions are validated on a monthly basis by SaharaAI.

Overview

The module is powered by advanced machine learning techniques, including a forest-based ensemble classifier approach, enabling the model to generate reliable predictions based on multiple factors. With users being able to quickly act on predictions via "Quick Buy" or "Quick Sell" directly from the predictions dashboard on xFractal's dApp, and you can also track PnL, accuracy, and daily prediction impact.

How It Works

Oblivia 1.0 aggregates and blends a variety of inputs including:

• On-chain activity

• Price trends & volatility

• Social metrics & sentiment

• Technical indicators (e.g. MA, RSI)

We use dimensionality reduction to ensure no single data point hogs the spotlight. The results are balanced and reliable predictions that thrive in Solana’s volatile environment.

These features are passed through multiple forest-based classifiers, each outputting a probability of upward price movement.

Model Ensamble

At the core of Oblivia 1.0 lies a multi-model ensemble architecture, which is characterized by robust decision boundaries and mitigates the limitations inherent in single-model approaches. The principal components of this architecture include:

• Forest-Based Classifiers Multiple instances of decision tree ensembles are trained on distinct subsets of the feature space. Each classifier outputs a posterior probability, Pi​, representing the estimated likelihood of an upward price movement given the observed features.

• Weighted Fusion Mechanism The module employs a weighted aggregation of individual classifier outputs. Letting wi​ denote the weight assigned to classifier i . The weighting coefficients wi​ are calibrated through cross-validation, optimizing for both precision and recall across varying market regimes. This method mitigates overfitting by smoothing idiosyncratic fluctuations present in any single model’s output. The ensemble probability score is computed as:

• Dynamic Thresholding The resulting ensemble probability Pensemble​ is continuously compared against a dynamically adjusted threshold θ . The threshold θ adapts to temporal variations in market volatility and liquidity conditions. The final binary decision is given by, where "1" implies an anticipated upward price movement and "0" indicates a potential downward or neutral trend:

If a model doesn’t reach 89%+ precision, no prediction is made.

Performance Metrics

Our main performance challenge is finding the right threshold—the minimum confidence required to issue a signal. There’s a natural trade-off:

• Higher threshold = higher precision (fewer false positives)

• But lower threshold = higher recall (we don’t miss the winners)

We've tuned our system to find a balance, increasing the precision meaningfully while maintaining enough recall to keep the predictions relevant and broadly usable.

This calibration ensures that the most confident predictions tend to be right—and that users can filter accordingly based on their strategy.

Correlation Between Threshold & Precision

One of the most informative takeaways from our evaluation:

• As we increase the classifier threshold, precision climbs consistently, but recall drops.

This relationship is shown clearly in the chart below. It confirms that confidence is predictive of correctness. Our optimised thresholds ensure traders have enough predictions and high accuracy in order to maximise their potential gains.

Market Cap Clustering

To better understand performance dynamics, we clustered tokens by market cap into three segments: Low (<$5m), Mid (<$50m), and High ($50m).

Here’s what we found:

• All three groups show similar overall accuracy, validating the model’s general robustness

• Mid MC tokens have the highest precision, suggesting they offer clearer price signals and less noise

• High MC tokens dominate on recall, likely due to more consistent behavior and richer data

• Low MC tokens don’t stand out—possibly due to excessive volatility, limited on-chain data, or social noise distortion

This opens the door to segment-specific tuning, which is already in the pipeline.

Clearly on the following area chart it is shown that the best performers are between Mid and High MC since their markets are usually more well established and less volatile (compared with Low MC tokens).

Future Developments

We’re currently working on a quant model to integrate directly into our predictions, therefore having a fully autonomous, real-time predictive execution stack. Once we have the quant strategy fully developed and merged with the prediction model, our predictions will become self-fulfilling, and we’ll be forced to only serve a small group of power users.

Regardless, there will be more spin-offs and products that will come out of this unified model.

And we're actively working on multiple improvements to increase both usability and predictive power:

• Market cap-specific models for optimized thresholds per segment

• Meta-learning systems to auto-tune model parameters based on live feedback

• Volatility-aware dynamic thresholds

• Token coverage expansion as data availability grows

• Visualization dashboards to explore predictions & performance in real-time on the dApp

Unlocking Predictions

Predictions can be unlocked for 1 XCC. Once unlocked, users can review the prediction output and instantly act on it by selecting Quick Buy.

Oblivia 1.0 is more than a signal engine—it’s a new way to engage with market intelligence on Solana.

Confident predictions. Transparent metrics. Real-world utility.