💻 computer science

Visual Chart Representations for Cryptocurrency Regime Prediction: A Systematic Deep Learning Study

This systematic deep learning study demonstrates that for cryptocurrency regime prediction, simple 4-layer CNNs trained on raw, low-resolution candlestick charts outperform complex pretrained architectures and advanced image encoding methods, achieving an AUC-ROC of 0.892 while benefiting significantly from ImageNet transfer learning.

Original authors: Dustin M. Haggett

Published 2026-05-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Dustin M. Haggett

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

Imagine you are trying to predict the weather. For centuries, sailors and farmers looked at the sky, the clouds, and the wind to guess if a storm was coming. They didn't need a supercomputer; they just needed to "see" the patterns.

This paper asks a similar question about the stock market and cryptocurrency: Can we teach a computer to "see" the market just like a human trader does?

Instead of feeding the computer raw numbers (like "Bitcoin is $50,000"), the researchers fed it pictures of the market—specifically, the colorful candlestick charts that traders look at every day. They wanted to see if a computer could look at these pictures and guess whether the market is about to go up (a "Bull" regime) or down (a "Bear" regime).

Here is what they discovered, broken down into simple stories:

1. The "Raw Photo" vs. The "Mathematical Art"

The researchers tried two ways to turn market data into pictures:

The Raw Photo: A standard candlestick chart (the red and green bars you see on news sites).
The Mathematical Art: A complex, abstract image created by turning the numbers into a grid of angles and sums (called Gramian Angular Fields).

The Result: The computer was terrible at the "Mathematical Art." It got confused. However, it was quite good at the Raw Photo.

The Analogy: It's like trying to identify a dog. If you show a child a real photo of a dog, they know it immediately. If you show them a complex mathematical equation that describes a dog, they have no idea what it is. The computer learned that the simple, familiar picture was the key.

2. Less Clutter is Better

Traders often add extra lines to their charts: volume bars, moving averages, and "Bollinger Bands" (which look like a safety net around the price). The researchers tested if adding these extra lines helped the computer.

The Result: Adding extra lines actually hurt the computer's performance. The simplest charts—just the price candles with no extra doodles—worked the best.

The Analogy: Imagine trying to read a street sign in a foggy city. If the sign is clean and simple, you read it easily. If someone paints 50 extra arrows, graffiti, and advertisements all over the sign, you can't see the direction anymore. The computer got distracted by the "visual clutter."

3. Small Brains Beat Big Brains

In the world of AI, bigger usually means better. The researchers tested a tiny, simple computer brain (a small neural network) against massive, famous ones (like ResNet and Vision Transformers) that are usually used to recognize cats and cars.

The Result: The tiny, simple brain won. The massive, complex brains actually failed because they tried to memorize the small amount of data they were given, rather than learning the actual patterns.

The Analogy: Think of a master chef (the big model) trying to cook a simple grilled cheese sandwich using a 50-page recipe. They overthink it and burn the bread. A home cook (the small model) with a simple recipe makes the perfect sandwich. Sometimes, you don't need a supercomputer; you just need a simple tool that fits the job.

4. The "Training Wheels" Worked

The researchers tried using "Transfer Learning." This is like taking a computer that was already trained to recognize millions of photos of dogs, cars, and trees (from the internet) and asking it to look at stock charts.

The Result: Even though stock charts look nothing like dogs or trees, the computer still learned faster and did better.

The Analogy: Imagine teaching a person who is already an expert at spotting shapes in clouds to spot shapes in a new game. Even though the clouds aren't the game, their ability to "see shapes" helps them learn the new game much faster. The computer used its "shape-finding" skills to understand the charts.

5. Bitcoin vs. The Rest

The computer was great at predicting Bitcoin, okay at Ethereum, and terrible at the S&P 500 (traditional stocks).

The Result: Bitcoin charts seemed to have clearer, more consistent patterns than the stock market.

The Analogy: Bitcoin is like a chaotic dance floor where everyone is moving to the same loud beat, making it easier to predict the next move. The stock market is like a quiet library where people are whispering different things; it's much harder to hear a pattern.

The "X-Ray Vision" (What the Computer Saw)

Finally, the researchers used a special tool called GradCAM to see where the computer was looking on the chart.

They found the computer focused on the most recent candles (the right side of the chart).
It paid extra attention to big, bold candles (days with huge price swings).
It noticed when the color changed from green to red (a trend reversal).

The Conclusion: The computer wasn't just guessing randomly. It was looking at the exact same things human traders look for: recent momentum and big price moves.

The Bottom Line

The paper concludes that simplicity wins. To predict market trends using pictures:

Use simple, raw candlestick charts (no extra indicators).
Use a small, simple computer model.
Don't overcomplicate the image resolution.
Expect better results on cryptocurrencies than on traditional stocks.

The authors warn that this is a starting point. They used a relatively small amount of data (about 500 samples), and real-world trading involves fees and risks that this study didn't fully test. But the core message is clear: sometimes, the simplest way to see the future is just to look at the picture clearly.

Technical Summary: Visual Chart Representations for Cryptocurrency Regime Prediction

Problem Formulation

This study addresses the underexplored application of deep learning to financial chart images for regime prediction. While technical traders have historically relied on visual analysis of candlestick patterns to forecast market movements, the systematic evaluation of how different visual representations and neural architectures perform on this task remains limited. The paper formulates regime prediction as a binary image classification task: given a chart image representing $N$ days of historical price data, the model must predict whether the asset will enter a "bull" regime (defined as a return exceeding a 2% threshold over a 7-day forward horizon) or a "bear" regime. The research seeks to determine optimal configurations for image encoding, chart components, model architecture, and transfer learning strategies across Bitcoin, Ethereum, and S&P 500 data.

Methodology

The authors conducted eight controlled experiments spanning 2018–2024, utilizing daily OHLCV data from Bitcoin (BTC-USD), Ethereum (ETH-USD), and the S&P 500 ETF (SPY). Each dataset contained 500 samples, split chronologically into 70% training, 15% validation, and 15% testing to prevent look-ahead bias.

Image Encoding and Components

Three image encoding methods were compared:

Raw Candlestick Charts: Standard visualizations rendered via mplfinance, encoding Open, High, Low, and Close prices.
Gramian Angular Fields (GAF): Mathematical encodings transforming time series into polar coordinate matrices (Summation and Difference variants).
Multi-Channel GAF: Stacking separate GAF matrices for Open, Range, and Close prices into RGB channels.

Five chart component configurations were evaluated, ranging from price-only candlesticks to complex overlays including volume bars, Simple Moving Averages (SMA), and Bollinger Bands.

Model Architectures

Four architectures were tested:

Simple CNN: A custom 4-layer convolutional network (~422k parameters).
ResNet18: A standard 11.2M parameter architecture.
EfficientNet-B0: A 4.0M parameter architecture.
Vision Transformer (ViT-Tiny): A 5.5M parameter attention-based model.

Experiments varied image resolution (64×64, 128×128, 224×224), lookback windows (14, 30, 60, 90 days), and the use of ImageNet transfer learning. Models were trained using AdamW with class weighting to handle regime imbalance, and performance was primarily evaluated using AUC-ROC to ensure threshold-independent discrimination capability.

Key Contributions

Systematic Comparison of Encodings: The first comprehensive comparison of raw candlestick charts versus mathematical encodings (GAF) for financial regime classification.
Ablation of Visual Components: A study quantifying the impact of adding technical indicators (volume, SMA, Bollinger Bands) to chart images.
Architecture Benchmarking: The first direct comparison of CNNs versus Vision Transformers on candlestick chart images.
Transfer Learning Analysis: An evaluation of the efficacy of ImageNet pretraining in the specialized domain of financial charts.
Interpretability: The application of GradCAM to visualize and validate the visual features CNNs learn from financial charts.
Open Source: Provision of code, trained models, and experimental results for reproducibility.

Experimental Results

Image Encoding

Contrary to the hypothesis that mathematical encodings preserve temporal dependencies better, raw candlestick charts significantly outperformed GAF variants. The raw candlestick approach achieved an AUC-ROC of 0.760, whereas GAF Summation and Multi-channel variants scored below random chance (0.310 and 0.252, respectively). The authors suggest that mathematical transformations may destroy the spatial patterns essential for CNN feature extraction.

Chart Components

Simplicity proved superior. Price-only candlesticks yielded the highest AUC-ROC (0.815). The addition of volume bars, moving averages, or Bollinger Bands consistently degraded performance, with the "All Indicators" configuration dropping to 0.704. This suggests that technical indicators either introduce visual noise or provide redundant information already implicit in the price data.

Lookback Window and Resolution

Lookback: A 30-day window achieved the best performance (0.734 AUC), while shorter (14-day) and longer (60/90-day) windows performed worse.
Resolution: A resolution of 128×128 pixels achieved the peak AUC-ROC of 0.892. Lower resolution (64×64) lost fine-grained patterns, while higher resolution (224×224) introduced noise or required more data than available.

Model Architecture

A simple 4-layer CNN (~422k parameters) outperformed all larger, pretrained models. It achieved an AUC-ROC of 0.831, surpassing ResNet18 (0.662), EfficientNet-B0 (0.365), and ViT-Tiny (0.570). The larger models exhibited severe overfitting on the limited 500-sample dataset, suggesting that for this specific domain and data volume, model complexity is detrimental.

Transfer Learning

Despite the domain gap between natural images and financial charts, ImageNet pretraining consistently improved performance. For ResNet18, pretraining increased AUC-ROC by 4.1 percentage points; for EfficientNet-B0, the gain was 16.1 percentage points. This indicates that low-level features (edges, textures) learned from ImageNet transfer effectively to chart analysis.

Cross-Asset Generalization

Performance varied significantly by asset. Bitcoin showed the strongest predictability (0.734 AUC), while Ethereum (0.492) and the S&P 500 (0.459) performed near or below random chance. The authors attribute Bitcoin's predictability to the dominance of retail traders relying on technical analysis, creating self-fulfilling patterns.

Interpretability

GradCAM visualizations confirmed that the model focuses on recent price action (the right side of the chart) and candles with large bodies (significant momentum). The model also attended to color transitions (trend reversals), aligning its learned features with established technical analysis principles.

Significance and Claims

The paper concludes that "simplicity wins" in visual regime prediction. The optimal configuration identified is a simple CNN trained on raw, price-only candlestick charts at 128×128 resolution with a 30-day lookback.

Key implications drawn by the authors include:

Data Representation: Raw visualizations preserve predictive information better than abstract mathematical encodings.
Feature Engineering: Adding technical indicators to images is counterproductive; less visual clutter leads to better model performance.
Model Selection: Practitioners should prioritize simple models over large, pretrained architectures when data is limited, as complex models overfit easily.
Transfer Learning: Even with a significant domain gap, ImageNet pretraining provides a measurable boost in performance.
Market Specificity: Visual patterns are more predictive in cryptocurrency markets (specifically Bitcoin) than in traditional equities, likely due to market microstructure and participant behavior.

The authors emphasize that these findings offer practical guidance for designing visual trading systems, suggesting that engineering efforts should focus on data quality and problem formulation rather than increasing model complexity. They caution that results may vary with larger datasets and that predictable patterns could be arbitraged away as markets evolve.