T1 - Analysing SPL Price Action
Collision Fund
2025
01/
Introduction
Collision Fund is proud to announce T1, a cutting-edge AI model designed for analyzing price action for SPL tokens in the Solana ecosystem. T1’s purpose is to provide traders with an unparalleled edge in navigating Solana’s fast-moving markets. By leveraging advanced machine learning on an enormous dataset, T1 can detect trading opportunities and risks that are impossible to spot with conventional tools. Here are some of T1’s unique advantages in the Solana ecosystem:
Unmatched Training Scope: T1 is trained on historical data from over 80,000 Solana-based tokens, covering virtually the entire ecosystem. This broad exposure means the model has seen countless market scenarios, from major coins to the smallest memecoins.
High-Resolution Data: The model processed 841 billion+ one-second candlesticks during training – an extremely granular view of price movements. This allows T1 to recognize subtle patterns and react to price changes in real time.
Real-Time Analysis: T1 operates on live data streams, enabling it to issue buy or sell signals within seconds. In a 24/7 crypto market that never sleeps, such instant analysis ensures traders can respond to volatility immediately, without waiting for human analysis. AI never needs a break, scanning markets continuously and objectively.
02/
Architecture
At the heart of T1 is a sophisticated AI architecture tailored for financial time-series analysis. The model is built on a Transformer Neural Network, a type of deep learning architecture that has revolutionized sequence analysis in recent years. Transformers were first popularized in natural language processing (they power large language models like GPT-3), but their strength at capturing long-range dependencies in sequences makes them equally well-suited to price data. In T1’s case, the Transformer is used to analyze sequences of candlesticks rather than words. This design allows the model to consider a wide context window of past price action when making decisions, rather than just the most recent data. Academic research and our internal benchmarks have shown that transformer-based models can outperform traditional time-series models in cryptocurrency price prediction by extracting complex temporal patterns. T1 leverages this capability to understand both short-term signals and longer-term trends on Solana.
Here’s a high-level look at how T1’s architecture works: The input to the model is a sequence of recent candlesticks for a given token (along with technical indicators derived from them). These sequences can be quite long – T1 can analyze many hours of second-by-second data in one go, if needed, to find patterns that develop over time. The Transformer component processes this sequence through multiple layers of “self-attention,” which is a mechanism that allows the model to weigh the importance of different time steps relative to each other. For example, if a price pattern forming now is similar to one that happened 10 minutes ago, the model’s attention mechanism can draw that connection and alert the decision-making layers. This architecture is adept at identifying intricate patterns – it’s the same principle that enables language models to understand context and nuance in a paragraph, applied here to streams of price and volume data. Thanks to the Transformer's design, T1 can capture both local patterns (like a sudden volume spike in the last 5 seconds) and global context (like a gradual upward trend over the past hour) within a unified framework.
On top of the Transformer core, we integrated a reinforcement learning (RL) module that fine-tunes T1’s outputs for trading objectives. While the Transformer excels at pattern recognition and prediction, the RL component is what turns those predictions into actionable trade decisions. We utilized reinforcement learning to teach T1 how to act on its insights – when to confidently issue a buy signal versus when to hold off, how to balance the risk of false signals, etc. In practice, this means T1 doesn’t just forecast price movements; it actively learns policies for profitable trading. The model was trained to maximize long-term rewards (like portfolio gains) rather than any short-term metric. This combination of a predictive model with a decision-making overlay is inspired by the way advanced AI systems (like AlphaGo or certain trading bots) operate, where they first evaluate the situation and then choose an optimal action. By blending supervised learning (for prediction) with reinforcement learning (for action optimization), T1’s architecture is designed to output high-conviction trading signals in real time.
It’s useful to contrast T1’s approach with traditional trading algorithms which do not work on low liquidity, high volatility tokens in the Solana ecosystem. Classic strategies often rely on fixed indicators (moving averages, RSI, etc.) or statistical models like ARIMA/GARCH that assume a certain statistical structure in price data. These methods have limited memory (for instance, a 50-day moving average only “remembers” 50 days of data) and cannot easily account for the kind of nonlinear, regime-shifting behaviors seen in crypto. In a rapidly evolving market, their performance can break down because they can’t adapt on the fly. T1’s Transformer-based architecture, however, has no such hard-coded limits – it can flexibly adjust to new patterns as they emerge. Moreover, GPT-style models (to draw an analogy) learn from raw data without human-imposed rules, often discovering patterns that humans weren’t aware of. T1 functions in a similar vein: it can pick up on signals across Solana markets that aren’t captured by any single indicator, essentially learning its own indicators from the data. For example, it might learn a complex combination of order book changes and volume bursts that reliably precedes a pump – something a human trader might only notice after years of experience, if at all.
Another advantage of T1’s AI core is its ability to generalize across tokens. Because it was trained on tens of thousands of tokens, the model built an abstract understanding of market behavior. It can apply that learned intuition to new tokens that appear, even if it’s never seen them before, akin to how a language model can understand a sentence with new words by context. Traditional models typically need to be configured per asset (a momentum strategy that works for SOL might not work for a tiny new token). In contrast, T1’s deep learning approach allows it to transfer knowledge from one asset to another. If a brand-new memecoin starts trading today, T1 can immediately analyze its price action in context of similar patterns it learned from other tokens. This is a profound shift from the one-size-per-asset approach of the past – it’s AI-powered adaptability that matches the dynamic nature of Solana’s markets.
03/
The Data Pipeline
Developing T1 required assembling one of the most comprehensive Solana market datasets ever. Collision Fund began by sourcing historical price data for over 80,000 Solana tokens. This included everything from SOL itself and major SPL tokens to tiny memecoins and LP tokens. Data was gathered from decentralized exchanges, on-chain order books, and other public blockchain data sources to ensure full coverage. Every trade, order, and price point was collected. However, raw trade data can be irregular and noisy – especially in a decentralized context – so the next step was structuring this data into uniform candlesticks.
We converted the trade history of each token into 1-second OHLC candlesticks (Open, High, Low, Close bars). This means for every second in the timeline, we have a summary of price movements and volume, even for tokens that traded sporadically. Handling data at one-second resolution across tens of thousands of assets was a massive undertaking, but it was crucial for capturing the rapid dynamics of low-cap tokens. Many price moves on Solana happen in a matter of seconds, so higher-level summaries (like 1-minute or 5-minute candles) might miss critical detail. By standardizing on 1-second intervals, we ensured T1 could learn from the smallest tremors in the market.
Data preprocessing was applied rigorously to maintain quality. We cleaned and normalized the data by filling in any gaps (for instance, if a token had no trades in a given second, that candle is marked with the last known price to avoid blanks) and filtering out erroneous data points. Any anomalous spikes caused by glitches or manipulation (e.g. a trade that momentarily reports an outlandish price due to a decimal error) were treated as outliers and removed or capped. This process follows best practices for AI data preparation: for example, filling missing values and removing random noise helps the model focus on meaningful trends. We also ensured all price data was consistently formatted (denominated in USD where applicable, volumes normalized, etc.) so that the model wouldn’t be thrown off by inconsistent units or scales.
To further improve data quality, we filtered out tokens with insufficient data. Many Solana tokens are created and then quickly abandoned, resulting in extremely sparse trading records. Including such tokens could introduce more noise than signal. We applied criteria to exclude tokens that didn’t meet minimal liquidity or longevity thresholds (for example, tokens that only traded for a few seconds or had virtually no volume after launch). By doing so, we focused T1’s training on tokens that had at least some meaningful trading activity, ensuring the model learns generalizable patterns rather than one-off anomalies. Throughout this pipeline, our team continuously validated the integrity of the data – cross-checking with multiple sources and manually inspecting samples. We recognize that AI trading bots rely heavily on accurate data, and that poor or inconsistent data leads to bad predictions. Thus, a tremendous amount of effort went into building a Solana price database that is clean, reliable, and comprehensive.
04/
Training Process & Infrastructure
Building T1’s intelligence required an immense training effort. We split the training process into multiple stages and used powerful infrastructure to ensure the model reached a high level of performance:
Supervised Learning on Historical Data: In the first phase, T1 was trained in a supervised manner using the historical dataset of candlesticks and curated trading signals. We created training examples from the data where the “input” was a sequence of market data (price movements, volumes, etc.), and the “label” was an ideal action to take (such as buy, sell, or hold) at that point, based on what happened next. These labels were derived from historical outcomes – for example, if a token’s price was about to surge by 100%, the data just before that move would be labeled as a buy opportunity in hindsight. By doing this for millions of past events, we taught T1 to recognize the precursors of major price increases or decreases. The model adjusted its internal parameters to reduce prediction errors, essentially learning to mimic successful trading decisions of the past. This phase exposed T1 to 841 billion+ candlesticks worth of scenarios, engraving into it the patterns that often lead or lag significant market moves. Training on such a vast corpus was crucial: deep neural networks benefit from huge amounts of data to capture subtle signals, especially in chaotic markets. Over this supervised training, T1 gradually improved its ability to forecast short-term price direction and identify setups that historically led to profitable trades.
Reinforcement Learning Fine-Tuning: While supervised learning gave T1 a strong predictive foundation, the next step was to fine-tune it for actual trading performance. We employed reinforcement learning, allowing the model to interact with a simulated trading environment and learn from the outcomes. In these simulations, T1 could autonomously decide to buy or sell tokens (using historical price sequences as the sandbox) and would then observe the result of its actions on portfolio returns. We defined a reward function tied to trading profitability – for instance, making a profitable trade would grant a positive reward, while a bad trade or missing a big opportunity could incur a penalty. T1 used these signals to adjust its decision-making policy. Through countless trial-and-error iterations, the model learned strategies that maximize cumulative rewards, effectively learning to trade by experience. This reinforcement phase is akin to training a game-playing AI: T1 “played” the game of trading over and over, improving each time. Importantly, reinforcement learning taught T1 to be output-focused – it wasn’t enough to predict correctly, it had to decide in a way that would yield profits in a real trading scenario. This process helped the model calibrate its confidence in signals (to avoid over-trading) and balance between exploitation (capitalizing on known patterns) and exploration (identifying new patterns). By the end of this phase, T1 had evolved from a pure predictor into a full-fledged trading agent that aims for the best long-term outcomes.
High-Performance GPU Infrastructure: The scale and complexity of training T1 demanded enterprise-grade computing power. Collision Fund utilized a cluster of 128 NVIDIA A100 GPUs (Graphics Processing Units) in parallel to run the training computations. The NVIDIA A100 is one of the world’s most advanced AI accelerators, featuring 80 GB of high-speed memory per GPU and exceptionally high throughput. In fact, the A100 offers over 2 terabytes per second of memory bandwidth – the fastest in the world – which enables it to handle enormous models and datasets efficiently. By distributing T1’s training across 128 of these GPUs simultaneously, we achieved the computational muscle needed to iterate over hundreds of billions of data points and billions of model parameters. Training a model of T1’s size on a single machine would take impractically long (years); on our GPU cluster, we were able to complete training in a matter of weeks. The infrastructure was configured with high-speed interconnects (InfiniBand networking) to allow the GPUs to communicate and share gradients quickly during training. We also implemented advanced techniques like mixed-precision training and gradient checkpointing to optimize memory usage and speed. This kind of setup – using dozens of top-tier GPUs in concert – is typically only seen in leading AI research labs. It’s similar to the hardware scale used for training frontier models like large language models, underscoring the ambition behind T1. By investing in state-of-the-art computing resources, we ensured that T1’s training was both thorough and efficient. The result is a model that encapsulates an unprecedented amount of market knowledge, attained through a heavy but necessary computational effort.
05/
Benchmarking
After training, we rigorously benchmarked T1 to measure its trading performance. The results have been striking. By our definitions and test scenarios, T1 achieves approximately 94.3% accuracy on buy signals and 81.3% accuracy on sell signals – far exceeding typical standards in algorithmic trading. It’s important to clarify how we define “accuracy” in this context. For buy signals, we count a signal as “accurate” if the token’s price rose significantly after the model’s recommended entry. Specifically, for micro-cap tokens (those with a market cap under $4 million at the time of signal), an accurate buy means the price later increased by +300% or more. For somewhat larger tokens (market caps under $180M), which tend to move less explosively, we set a lower threshold of +45% price increase to deem the buy signal successful. These are very high bars – we’re not talking about just a 5% or 10% uptick, but substantial rallies. The fact that T1’s buy signals hit these marks 94.3% of the time is a testament to the model’s ability to pinpoint true breakouts. In other words, when T1 says “buy,” the odds overwhelmingly favor that the asset will soon surge by a notable amount.
On the flip side, T1’s sell signals have demonstrated 81.3% accuracy according to our tests. For sell signals, accuracy means that after the model issued a sell (or avoid) recommendation on a token, the price subsequently dropped sharply. We observed that in over 81% of such cases, the token’s price fell by more than 50% not long after the signal, vindicating T1’s warning. This level of performance indicates that T1 is adept not only at finding bullish opportunities, but also at flagging overvalued or vulnerable situations before major downturns occur. Essentially, it often manages to get traders out before a coin “rug pulls” or collapses under profit-taking. Capturing over half of such catastrophic drops in advance is a remarkable feat given how quickly crashes can happen in crypto.
To put these numbers in perspective: a conventional algorithmic strategy in crypto might consider a 60-70% win rate to be excellent. In fact, a recent study using a neural network for Bitcoin trading achieved about 66% accuracy on its predictions and was deemed profitable under various conditions. T1’s ~94% buy accuracy under stringent criteria represents a significant leap in precision. Achieving above 90% accuracy in any trading prediction task is extremely challenging – markets have a way of humbling models with noise and random events. T1’s performance suggests that the combination of its vast training data and advanced architecture is capturing something very fundamental about Solana price movements. It’s finding real alpha in places where other algorithms largely produce coin flips. Moreover, because our accuracy definition requires large price moves, a 94.3% accuracy implies a very low false-positive rate. T1 isn’t throwing out a bunch of signals and hoping some stick; it’s selectively issuing alerts that almost always correspond to big moves, thus filtering out the false alarms that plague many trading systems.
We also evaluated T1 across different market conditions and volatility regimes to ensure its performance is robust. The crypto market goes through cycles – bull markets where everything is rallying, bear markets where liquidity dries up, and sideways or choppy periods. We tested T1 on historical data segments representing all these scenarios. Impressively, the model maintained high accuracy throughout. During bullish conditions, there are more opportunities and T1 naturally produces more signals, but we found its precision remained very high even as the market got crowded with potential trades. In bearish or high-volatility conditions, where many assets are crashing or whipsawing, T1 became more conservative (fewer signals) but still upheld a strong success rate on the signals it did produce. This adaptability is crucial – many trading bots that work in trending markets fail in choppy ones (and vice versa). T1’s blend of machine learning models equips it to handle both. This aligns with findings in academic research that emphasize evaluating model robustness under various market scenarios. We’ve essentially stress-tested T1 in different seasons of the market, and it consistently identified profitable plays or timely exits.
Another aspect of performance is latency and real-time reliability. In live paper-trading simulations, T1 was able to ingest real-time Solana blockchain data and output signals with only a second or two of delay. This means that even during sudden spikes in volatility, the model kept up with the flow of information and didn’t lag on giving an alert. For traders, this is vital – an accurate signal delivered too late is of little use. We designed T1 from the ground up for low-latency operation, and our benchmarks confirmed that it meets the real-time requirements of high-frequency crypto trading.
It’s worth noting that while the headline accuracy figures are extremely high, we treat T1’s signals with appropriate caution. No model is perfect, and the remaining ~6% of buy signals that didn’t meet the success threshold (or ~19% of sell signals that didn’t materialize into big drops) remind us that uncertainty is inherent in markets. In some cases, a token might only rise 30% instead of the hoped-for 60%, or it might hit the target gain but then abruptly reverse due to external news. We analyze every “miss” to continually refine the model and our usage of it. Often, those misses still result in small gains or flat outcomes, not catastrophic losses, because the model tends to avoid truly bad trades. Nonetheless, understanding the few failures is key to continuous improvement.