3D Wallet Explorer
This page explains how the 3D Wallet explorer was built, how to navigate it, and how to interpret what you see.
Module summary
The 3D Wallet Explorer shows the entire trading network as a map in three dimensions.
Each point represents a wallet, and each color represents a group (or “cluster”) of wallets that behave in similar ways.
Hoku uses machine learning to place wallets that share similar trading patterns close to each other in space. These patterns include basic wallet characteristics—such as how old a wallet is and how many trades it has made—as well as more advanced factors that describe a trader’s style and risk level.
By looking at this view, you can:
Explore relationships between traders. Nearby wallets usually have similar behaviors or strategies.
Click any wallet to see more information, such as its balances, trading history, and profit or loss.
View cluster summaries to understand how that group of traders performs overall—how profitable or risky they are compared to others.
Search with DeepLeap, a natural-language feature that lets you type questions like “find wallets that made 10x buying meme tokens early.” Hoku will search the database and highlight matching wallets.
This feature helps you see patterns that are hard to spot in tables or lists. For example, it makes finding similar traders far easier than traditional explorers. You can quickly identify successful trading groups, discover interesting strategies, and better understand your own position within the wider market.
Features
Network Toggle
The Network Toggle lets you switch between different blockchain networks instantly. You’ll find a simple selector that controls which network’s data you’re viewing at the top of the window.
Hōkū is currently live on Base, Ethereum, and BNB Smart Chain. When you change the network, every part of your dashboard — your portfolio, cluster position, trending tokens, and personalized recommendations — updates automatically to reflect your activity and relationships within that specific chain.
This makes it easy to monitor how your trading style and performance vary across networks, or to explore entirely different ecosystems with a single click. If you use multiple chains, switching networks ensures this module always shows accurate, real-time insights for the one you’re focused on.
All modules in Hōkū are network-specific except the Trade Signal Alerts page. That means you'll need to select a specific network to view its data on all other modules where this toggle is present
Navigating the 3D Space
The 3D Wallet Explorer is designed to feel familiar and intuitive, similar to navigating in Google Earth. You can move freely through the map to explore millions of active wallets on a given blockchain network.
How to move around
Pan: Click and drag to move your view across the map.
Zoom: Scroll or pinch to move closer or farther away from the scene.
As you zoom in, more detail will appear. At wide zoom levels, you’ll see clusters and high-level patterns across the network. As you move closer, individual wallets begin to appear, revealing up to five million points of activity in total.
Exploring wallets and clusters
Click any wallet to open two panes on the right side:
Wallet Details: Shows specific information about the selected wallet, including balances, trades, and profit.
Cluster Overview: Summarizes the behavior and performance of the cluster that wallet belongs to, including showing a preview of the top trending tokens in that cluster (for more info on these trends see Trading Clusters).
From the Cluster Overview pane, you can also navigate between clusters using the next and previous buttons.
The camera will automatically reposition to the center of each cluster as you move through them, allowing you to explore the network one group at a time.
Control Buttons
On the Cluster Explorer specifically you'll find a series of buttons across the top.
From right to left, these buttons are:
Reposition camera to center
Search for a wallet with its address hash
Locate the connected wallet in the cluster explorer
Open DeepLeap, our AI-powered natural language search
DeepLeap Natural Language Search
DeepLeap executes natural langauge search queries against Deep3's research database allowing users to directly search for wallets (and soon tokens) based on a range of criteria. It continues to get smarter and more capable as our research database grows so if it's unable to return results today, check back again soon.
Within the DeepLeap menu, you'll find a familiar AI-chat interface and controls to search through the lists of addresses that it discovers (additional information on using DeepLeap is below).
Once you've executed a search, use the controls above the chat interface. They will allow you to (from left to right): hide the interface, return to the first wallet in the list, go back one wallet in the list and go forward to the next wallet in the list.
When you move from wallet to wallet in the list, an informational pane will open in the bottom right corner of the window.
Wallet Info Pane
The Wallet Info pane appears when you click on any wallet in the 3D Explorer. It gives you a detailed, easy-to-read overview of that trader’s recent activity and style.
Main elements
Wallet identity: Shows the wallet address, ENS name (if available), and profile image or PFP. You can copy the wallet address or add the wallet to your favorites.
Balance chart: Displays how the wallet’s total balance has changed over the past four weeks, so you can see short-term trends in value.
Current Balance: Shown in the top-right corner, this is the wallet’s total estimated USD value at the latest block.
Current Profit & Loss: Shown just under Current Balance, this is the wallet's total estimated profit or loss in USD terms.
Deep3 Proprietary Model Score Badges
Some wallets may include additional model-based scores generated by one of Deep3’s proprietary AI/ML products.
HODL-C1 Score: Predicts the likelihood that this trader behaves like a long-term holder. Scores closer to 100 suggest more consistent, patient holding behavior. Lower scores suggest short-term or high-turnover trading styles. For more information on this model see: HODL-C1 Model Details.
Predicted Win Rate: Estimates how often this trader’s past trades were profitable, shown as a percentage between 0% and 100%.
These scores give you a quick way to understand both the style (how they trade) and the quality (how effective they are) of a trader before you explore their activity in detail. They are only displayed as available.
Data Tabs
Just below the chart, you'll find 3 tabs that show additional data associated with the selected wallet.
Balances – Lists all tokens held by the wallet, including each token’s price, amount, and total USD value. You can also view 3 additional historical snapshots of the user's balances from 7, 14 and 21 days ago.
Trades – Shows recent token trades made by the wallet, helping you understand what assets the trader is currently active in.
P&L (Profit and Loss) – Displays gains or losses for each asset that the wallet has traded recently.
Cluster Info Pane
The Cluster Info pane appears next to the Wallet Info pane whenever you select a wallet in the 3D Explorer. It provides a broader view of the group—or cluster—that the wallet belongs to, helping you understand how that trader compares to others with similar activity.
Cluster Overview
At the top of the pane, you’ll see:
Cluster title & navigation – Identifies the cluster (for example, “Trading Cluster 3”) and enables "hopping" from cluster to cluster.
Cluster description – A short summary written by Hoku’s models describing the typical behavior of traders in this group. This may include the cluster’s size, how risky its traders tend to be, how often they take profits, or how long they hold their positions.
Cluster Metrics
Below the overview, three gauges summarize the cluster’s overall characteristics. Each score ranges from 0 to 100, where higher numbers indicate stronger presence of that trait.
Holding Score – How likely traders in this cluster are to hold tokens for long periods rather than trading frequently.
Risk Score – The average level of risk that these traders take based on their trading patterns and asset choices. Higher values suggest more aggressive strategies.
Profit Score – The average profitability of wallets in the cluster, showing how successful the group tends to be over time.
Trending Section
The Trending section highlights which tokens or contracts are currently most active among members of this cluster. This is useful for spotting early movements and understanding where traders with similar profiles are focusing their attention. It is a subset of the information shown on the Trading Clusters module screen.
You can filter the view using dropdown menus:
Transaction Type – Choose whether to show tokens being bought or sold.
Source – Filter to include transactions from only DEXs or all activity across the network.
Stable Coins – Show or hide stablecoin activity.
Time Range – Change the lookback window between 1 hr and 6hrs to see short-term vs. longer-term trends.
Each bar shows how popular that token is within the selected period, with the percentage representing the share of total transactions it was involved in across all wallets in this cluster during the selected time period. In the below example, 3.27% of all buy transactions in the past 1 hour within Trading Cluster 3 involved the VFY token.
Usage tips
This is the most advanced network explorer in the industry, so we're sure you'll find new and exciting ways to use it. Here's just a few to get you started.
Find wallets that trade like your favorite wallet
Most blockchain explorers only show wallets that have interacted with a given address — for example, wallets that sent or received tokens from it. But that doesn’t mean those traders actually behave alike. Our 3D Wallet Explorer takes this further by showing wallets that act alike, not just transact alike.
If you follow a wallet you respect or use for trading inspiration, simply enter it in the search bar. The explorer will automatically center on that wallet, and all the points nearby represent others with similar trading patterns, risk profiles, and holding behavior.
This makes it easy to discover dozens of wallets that mirror the same strategy — a quick way to uncover look-alike traders worth studying.
Use DeepLeap to uncover unique or “niche” traders
DeepLeap is a natural-language search tool that helps you find specific kinds of traders across millions of wallets — instantly. Want to find wallets that specialize in AI tokens, consistently buy meme coins early, or have made steady profits in low-cap markets? Just describe what you’re looking for in plain English.
DeepLeap searches the behavioral database and returns matching wallets or clusters that fit your description. It’s continually learning from new queries and trading data, so over time it supports more nuanced and precise requests — making it easier to discover profitable niches or emerging market behaviors without writing a single query.
Get trading inspiration from the network itself
If you’re the type of trader who loves exploring wallets on Etherscan or scanning DeFi dashboards for ideas, the 3D Wallet Explorer gives you a much more powerful way to do it.
As you click through wallets, you can see how each trader fits into the broader market structure. When something catches your eye, like a profitable cluster, an unusual trading rhythm, or a wallet with consistent success, you can easily explore the surrounding wallets to find others following the same approach. But, if you want a completely new perspective, jump to a different cluster or region to explore a totally different trading ecosystem.
Because the explorer organizes every active trader using comprehensive data and advanced algorithms, you’re no longer guessing where to look — you’re navigating the network of profitable behavior itself.
FAQ
How it works
Behind the smooth 3D interface of Hōkū’s Wallet Explorer lies a powerful machine learning system that processes millions of on-chain records to find patterns invisible to the human eye. Every point you see in the 3D map represents a trader whose behavior has been analyzed, quantified, and embedded in a multidimensional space.
This section explains how Hōkū builds that map in simple terms, from raw blockchain data to the final 3D view. It's written for a moderately technical audience but explained in simple terms.
Step 1: Gathering the data
Every month, Hōkū collects the complete trading history of wallets that have been active during the last 30 days. This process typically includes around 5 million wallets and over 100 million individual trades across the network.
For each wallet, the system measures around 20 features, such as:
Wallet age
Total number of trades
Total profit or loss
Average time between trades
This information helps the model understand each trader’s style and activity level.
Step 2: Build wallet behavior and risk features
For each wallet we compute two kinds of signals and treat them as one combined description of behavior.
Wallet characteristics (just a few examples)
We compute over 20 different descriptive features of each wallet, which is constantly growing and evolving as our research progresses.
Age: how long the address has been active.
Activity: number of trades, number of distinct tokens, trading frequency.
Capital & concentration: balances, size of typical trade, diversification.
Performance: realized profit/loss over time, volatility of outcomes.
Contextual risk features
Past trades are also summarized by how crowded or isolated those trades were at the time:
If a token was traded by many other wallets in the same period, that action is treated as lower risk (more consensus, more liquidity).
If a token was traded by few wallets in that period, it is treated as higher risk (niche, less consensus, thinner liquidity).
These signals are aggregated for each wallet (for example, “share of trades in high-risk contexts” or “typical crowd size when entering positions”). Together with the characteristics above, they form a single feature vector—a compact but expressive fingerprint of that wallet’s behavior.
Step 3: Generate embeddings
Raw features live in a high-dimensional table and many features overlap. We convert them into embeddings—dense numerical coordinates that capture the most informative structure.
PCA / SVD (linear compression) These methods find the main axes of variation (like rotating a cloud of points so that most of the spread lies along a few directions). They remove noise, reduce redundancy, and make the next steps faster and more stable.
UMAP (non-linear structure) Markets are not purely linear. UMAP preserves local neighborhoods—wallets that act alike stay close—while allowing curved, non-linear relationships. It is well-suited for behavior data where multiple strategies can coexist.
In practice, we use PCA/SVD to create a clean base representation, then learn a UMAP embedding on top of that representation. The result is an embedding space where “near” means “similar behavior.”
Step 4: Normalize the embeddings
Before comparing wallets or clustering them, we standardize each embedding dimension so that it has a consistent mean and variance.
Without normalization, a dimension with larger numeric range would dominate distance calculations.
With normalization, each dimension contributes fairly, so similarity reflects pattern, not scale.
Step 5: Find communities with HDBSCAN
Now we detect communities—areas where wallets are packed together in the embedding.
What HDBSCAN looks for
Imagine turning down the volume on sparse areas and listening for dense pockets of points. Those pockets are clusters. Areas without enough density are left as noise (unassigned) rather than forcing a poor label.
Why HDBSCAN fits trading data
Clusters can have irregular shapes (momentum chasers vs. liquidity farmers do not form tidy spheres).
Clusters can be very different sizes (a large class of conservative holders vs. a tiny pocket of niche risk-takers).
There are always edge cases; leaving them as noise is better than mislabeling them.
Key settings explained
min_cluster_size: “How big must a crowd be before we call it a community?”min_samples: “How strict should we be about density?” Higher values demand thicker neighborhoods and reduce accidental groupings.
Outputs you feel in the UI
A cluster label for most wallets → becomes the color you see.
Some wallets marked as noise → handled gracefully in the next step.
Step 6: Project the clustered space into 3D
We already used UMAP to shape the high-dimensional embedding. We use it again to project down to 3D for an interactive scene.
Goal: preserve local neighborhoods so that “near on screen” ≈ “similar in behavior.”
Trade-off: 3D cannot hold every nuance of the original space; the focus is on keeping nearby relations accurate and spreading clusters enough to be readable.
Stability: parameters are tuned for a stable layout so the scene feels consistent as you zoom and pan.
Step 7: Handle the "noise"
After projection, we refine the 3D space so it’s both visually clean and mathematically consistent. Two GPU processes handle this: a voxel-density filter that removes outliers and a nearest-neighbor placement that restores isolated points.
Voxel-density cleaning
To measure how tightly wallets cluster together, the 3D space is divided into thousands of small cubes called voxels—the 3D equivalent of pixels. Each voxel counts how many wallets fall inside it.
A brief Gaussian smoothing spreads those counts slightly so local variations look natural rather than noisy. We then estimate each wallet’s local density by sampling nearby voxels. Wallets in extremely empty areas (the lowest ≈ 3 %) are dropped from view. This keeps the map focused on meaningful, well-supported regions while preserving the correct wallet index.
Nearest-neighbor placement with FAISS
Some wallets are labeled as noise by HDBSCAN because their behavior doesn’t clearly fit any cluster.
We place these using FAISS ( Facebook AI Similarity Search )—a high-speed GPU library that finds the closest vector match among millions of points. Each unclustered wallet searches for its single most similar neighbor (k = 1) in the embedding space and adopts that neighbor’s 3D position and cluster color for display. This fills visual gaps and gives every wallet context without falsely assigning it full cluster membership.
Some extra detail on the three core methods
While each of the statistical methods we use has a long and well-understood history of being applied in a variety of different contexts and applications, it's very new to crypto. So here's a bit more explanation of why these approaches work in Hōkū.
UMAP: How it preserves neighborhoods
UMAP builds a graph where each wallet connects to its closest neighbors. It then finds coordinates where connected wallets stay close and loosely connected wallets drift apart. You control how local or global the layout feels by adjusting n_neighbors. Smaller values emphasize fine-grained strategy pockets; larger values show broader structure.
Strengths: excellent at keeping “who is like whom” intact, handles curved/complex shapes, works well at very large scale.
Limitations: axes are abstract; global distances can be compressed or stretched to keep neighborhoods faithful.
HDBSCAN: How it finds natural communities
HDBSCAN gradually changes the density threshold and watches where points stay grouped together. Groups that remain stable across thresholds are promoted to clusters. Points that do not consistently belong anywhere become noise.
Strengths: no need to pre-choose the number of clusters; resilient to odd shapes and size imbalance; honest handling of outliers.
Limitations: very small or very diffuse groups can be left as noise; results depend on how you define “dense enough” (min_cluster_size, min_samples).
FAISS: Why it is the right tool at this scale
Nearest-neighbor search in millions of vectors is hard to do quickly. FAISS uses optimized indexes (flat, IVF, HNSW, etc.) and GPU acceleration to return the closest vectors fast. For our use case we need accuracy (k=1 must be truly nearest) and speed (interactive UI), which FAISS provides.
Strengths: industrial-grade speed, exact or approximate options, GPU support.
Limitations: requires careful choice of index type for the size/latency trade-off; we tune this to keep UI interactions smooth.
Interpreting the final map
Distance communicates similarity of trading behavior and risk style.
Color is the community found by HDBSCAN.
Unusual wallets were visually anchored via FAISS but may not fit their neighborhood perfectly.
Trends in clusters (e.g., the “Trending” list) summarize what many similar traders are doing now; they are signals, not guarantees.
Last updated