💻 computer science

Time Series Forecasting based on Solana Digital Asset Dataset

This paper introduces the first comprehensive Solana digital asset time series dataset spanning March 2024 to March 2025, utilizing it to characterize ecosystem-wide market dynamics driven by DEX activity and SOL price while validating its predictive utility through state-of-the-art forecasting models like PatchTST.

Original authors: Yufeng Xiao, Minxing Wang, Pavel Braslavski, Dmitry I. Ignatov

Published 2026-08-21

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yufeng Xiao, Minxing Wang, Pavel Braslavski, Dmitry I. Ignatov

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the vast, digital economy of cryptocurrency, value does not move through banks or vaults but across public ledgers that record every transaction with precise timing. This world is dominated by two distinct types of activity: the trading of major assets on centralized exchanges, which function like traditional stock markets, and the frenetic, decentralized trading that happens directly on blockchain networks. One such network, Solana, has emerged as a high-speed highway for these digital assets, capable of processing thousands of transactions per second with minimal cost. Unlike older systems where data can be opaque or delayed, Solana's architecture leaves a complete, timestamped trail of every token's journey, from its creation to its final trade. Understanding how these digital tokens rise and fall in value is a challenge that has long puzzled analysts, largely because the market is driven by a chaotic mix of individual trader behavior, sudden news events, and the overall health of the network itself. The question researchers have long asked is whether it is possible to predict the future price of these assets by looking at the patterns of the network they inhabit, rather than just the history of the asset itself.

A team of researchers has taken a significant step toward answering this by constructing the first comprehensive dataset specifically designed to track the daily life of Solana tokens. They gathered information on 1,584 different tokens, observing them every day from March 2024 to March 2025. This collection is not merely a list of prices; it is a rich tapestry of 27 different variables for each token, including how many people are buying or selling, the total volume of trades, the amount of money sitting in liquidity pools, and the price of the underlying Solana network currency. By weaving together these token-level details with broader ecosystem signals, the researchers created a window into a market that is growing rapidly and changing quickly. Their goal was not just to build a better prediction tool, but to understand the actual behavior of this digital ecosystem during a period of intense expansion.

The analysis of this data revealed that the market does not move in isolation. The researchers identified two distinct moments of synchronized activity where the entire ecosystem seemed to breathe in unison: one in mid-November 2024 and another in mid-January 2025. During these peaks, trading volumes, the number of active wallets, and the number of new traders all surged simultaneously across the network. The second peak, occurring in January 2025, coincided with the launch of a specific token associated with a former U.S. president, an event that triggered a massive wave of speculation. The data showed a clear pattern of behavior leading up to this event: investors were accumulating liquidity, pouring money into the pools, only to begin withdrawing their funds shortly after the peak. This suggests that the fate of individual tokens is deeply tied to the broader mood of the Solana market and the activity levels of its decentralized exchanges, rather than existing as separate, isolated entities.

To test whether this detailed data could actually help predict the future, the researchers ran a series of forecasting experiments. They asked various computer models to predict the market value of these tokens three days into the future. The results showed that while simple statistical models that rely on repeating patterns could work well for some stable tokens, they struggled with the wild fluctuations of this ecosystem. The most successful approach was a deep learning model called PatchTST, which excelled at identifying complex, non-repeating trends in the data. A model based on a pre-trained language framework, when carefully adjusted to the specific data, performed nearly as well. These findings suggest that the dataset contains genuine, short-term signals that can be used to forecast market movements, provided the model is sophisticated enough to handle the volatility.

Further investigation into how these models made their decisions revealed what truly drives the predictions. The most important factors were not just the history of a single token, but the broader health of the network. The price of Solana itself, its moving averages, the total volume of trades across the entire decentralized exchange system, and the number of new trading pairs being created were all critical indicators. This confirms that in this digital economy, the price of a specific token is heavily influenced by the collective behavior of the entire market. The study concludes that by combining detailed token data with ecosystem-wide signals, it is possible to build a clearer picture of market dynamics. This work provides a new foundation for understanding how digital assets behave, offering a way to distinguish between random noise and meaningful market shifts in an environment that is often difficult to navigate.

Technical Summary: Time Series Forecasting based on Solana Digital Asset Dataset

Problem Statement
Accurate forecasting of digital assets on the Solana blockchain presents unique challenges due to the ecosystem's high volatility, rapid regime shifts, and the prevalence of "cold start" scenarios where newly launched tokens lack historical data. Traditional time series forecasting methods often struggle with the non-stationary nature of crypto prices and the scarcity of long-term asset-specific data. Furthermore, existing research frequently relies on centralized exchange (CEX) data or Bitcoin/Ethereum-specific features, neglecting the distinct microstructure of Solana's Decentralized Exchange (DEX) environment and its rich on-chain telemetry. There is a lack of curated datasets that simultaneously capture token-level behavior (transactions, liquidity) and ecosystem-level DEX activity to analyze how broader market sentiment propagates to individual assets.

Methodology
The study introduces a novel, curated dataset and employs a comparative forecasting framework to validate its signal content.

Dataset Construction: The authors compiled a time series dataset comprising 1,584 Solana tokens observed daily from March 24, 2024, to March 16, 2025. The dataset includes 27 variables per token, integrating:
- Token-level metrics: Transaction counts, active wallet counts, trading volumes, price indicators (open, close, high, low, average), and liquidity pool balances.
- Ecosystem-level metrics: Total Solana DEX volume, DEX trader counts, newly created trading pairs, and Solana (SOL) price indicators (including 50, 100, and 200-day moving averages).
- Filtering: Tokens were selected based on activity criteria (e.g., >2,000 unique holders, >1,000 daily transactions, >$20M total volume) and cleaned to remove missing values or insufficient observation windows.
Forecasting Task: The objective was to predict token market capitalization (price × circulating supply) three days ahead ( $H=3$ ). Market cap was chosen over raw price to normalize heterogeneous supply dynamics (issuance, burns, unlocks).
Model Evaluation: Eight models were benchmarked across three categories:
- Deep Learning: PatchTST (patch-based Transformer), Temporal Fusion Transformer (TFT), and TiDE (Time-series Dense Encoder). These were trained in a univariate manner (one model per token).
- Zero-Shot Approaches: Chronos (evaluated in both fine-tuned and zero-shot modes) and TimeGPT. These were applied in a multivariate/transfer manner (single model for all tokens).
- Statistical Baselines: Facebook Prophet and SeasonalNaive.
- Validation: A rolling window backtesting strategy was used for training, with a fixed out-of-sample test set (March 14–16, 2025). Performance was evaluated using MAE, RMSE, and MAPE. Statistical significance was assessed via Friedman tests, Wilcoxon signed-rank tests, and parametric linear mixed-effects likelihood-ratio tests.
Feature Importance: The study utilized permutation-based importance for TFT and SHAP values for TimeGPT to identify key covariates.

Key Results

Market Dynamics: Descriptive analysis revealed synchronized peaks in DEX volume, token trading volume, active wallets, and SOL price in mid-November 2024 and mid-January 2025. The January 2025 peak coincided with the launch of the "Trump" token, accompanied by a visible shift from liquidity accumulation to withdrawals, suggesting strong coupling between individual token dynamics and broader Solana market sentiment.
Forecasting Performance:
- PatchTST achieved the best overall rank (mean rank ≈2.97) across MAE, RMSE, and MAPE metrics, demonstrating superior robustness in capturing non-periodic fluctuations.
- ChronosFineTuned followed closely, significantly outperforming its zero-shot counterpart, indicating that asset-specific adaptation yields substantial gains in this regime.
- Statistical Baselines: Prophet and SeasonalNaive performed exceptionally well for many tokens, suggesting that a significant portion of the dataset retains short-term persistence, trend, or seasonal structures. PatchTST outperformed Prophet on 93 tokens and SeasonalNaive on 772 tokens, highlighting its ability to capture complex, non-periodic features missed by statistical models.
- TimeGPT exhibited the lowest variance in stability but the highest error rates overall, capturing the least additional non-periodic features compared to other models.
Feature Importance: Analysis confirmed that while token-specific price history is relevant, macro-level variables are highly informative. The most influential covariates included SOL price, SOL moving averages, total DEX volume, DEX trader counts, and the number of newly created pairs.

Significance and Contributions
The paper's primary contribution is the creation of the first Solana digital asset time series dataset designed specifically for forecasting and market-structure analysis. By linking token-level microstructure with ecosystem-level DEX activity, the dataset enables the study of how broader market conditions influence individual asset volatility.

The study demonstrates that:

Ecosystem Coupling: Individual token dynamics on Solana are strongly coupled to broader DEX activity and SOL price movements, particularly during periods of rapid growth and speculative events.
Model Viability: While deep learning models like PatchTST excel at capturing complex, non-periodic patterns, statistical baselines remain competitive for trend-dominated tokens. Zero-shot models show promise but benefit significantly from fine-tuning in this specific domain.
Data Utility: The dataset contains sufficient signal for short-term (3-day) market capitalization forecasting, validating the use of on-chain and DEX microstructure data as predictive features.

The authors explicitly state that the work is an empirical analysis and data-driven characterization of the ecosystem, refraining from trading recommendations or deployment advice due to the ethical risks of speculative misuse. The findings are presented as Solana-specific evidence, with cross-chain generalization left for future research.

More like this