Trade Signal Alerts
Module summary
The Trade Signal Alerts module delivers real-time notifications whenever Hōkū’s machine learning systems detect statistically significant trader-driven momentum across multiple clusters.
Each alert originates from the same live data that powers the Trending Tokens module, but instead of simply showing what’s trending, it predicts what’s likely to happen next.
Behind each signal are dozens of proprietary models that continuously analyze patterns in wallet behavior across the network. When a token begins trending in at least two clusters, the system evaluates whether the types of traders buying it match patterns that have historically preceded profitable moves. If they do, a trade signal alert is issued automatically.
There are two key aspects of these Trade Signal Alerts to keep in mind
Each alert includes predictions for three price targets across three time horizons, resulting in 9 separate predictions.
All alerts are recalculated continuously, so the prediction strength and price targets shown are always the latest prediction for the stated time horizon. Said another way:
Each trade signal is a rolling prediction, meaning its time horizons are always measured forward from the moment you view it, not from when it was first created.
What makes these alerts unique is that they don’t rely on traditional price indicators like RSI, moving averages, or Bollinger Bands. Instead, they forecast human behavior based on deep analysis of trading clusters, wallet history, and token-level network activity. While price movements themselves are notoriously unpredictable, the collective behavior of traders is measurable and often precedes visible market shifts.
By using this view, you can:
Catch early momentum – Receive alerts before market-wide indicators confirm a move.
See predictive context – Understand why a signal was triggered, including which trader groups are driving it.
Prioritize opportunities – Filter by probability, time horizon, or confidence level to focus on the most relevant trades for your strategy.
Integrate with your portfolio – Cross-reference alerts with your favorite wallets or cluster trends to verify whether smart money is acting.
This module transforms the collective behavior of millions of traders into actionable trade intelligence, giving you predictive, behavior-based insights far faster than conventional technical analysis ever could.
Features
Prediction window selector
The Prediction Window determines which alerts are displayed on the screen.
Each trade signal in Hōkū includes forecasts for 3 future time horizons: 1 hour, 4 hours, and 24 hours. Each signal shows the model’s expected price movement over that specific period starting from now.
When you select a window (for example, Now + 4 hours), you’ll only see alerts that include a predicted price increase over that exact timeframe. But, if a token has a predicted price change only for the 24-hour horizon and not for 1-hour or 4-hour intervals, it will appear only when the Now + 24 hours option is selected.
Predictions are rolling forecasts, not static snapshots.
Even if an alert was generated several hours ago, its time horizon (e.g., “4 hours”) still represents the model’s projection for the next 4 hours from the moment you view it, not from when the alert first appeared.
This ensures every alert you see corresponds precisely to the prediction window you’ve chosen. It also keeps the display focused on the time periods most relevant to your current trading horizon, whether you’re monitoring short-term volatility or longer trend development.
Active alerts
The Active Alerts section lists all trade signals that are still considered live by Hōkū’s predictive engine.
An alert remains active as long as its associated token continues to trend across at least two or more trading clusters, meaning that multiple distinct groups of traders are still showing coordinated interest in it.
Each row represents one currently active signal and summarizes key details at a glance:
Chain: the blockchain network where the signal originated (Base
, Ethereum 
, or BNB Smart Chain 
). Type: whether the system is predicting a Buy
or Sell 
momentum shift.Strength: the confidence level of the prediction (Low
, Medium 
, or High 
).Price Target: the model’s projected directional change for the selected prediction window (for example, “+50%” or “–5%”). The price thresholds predicted vary for low and high liquidity tokens.
Token: the token that triggered the signal.
For (Clusters): the specific trading clusters driving the signal. These indicate which groups of wallets are responsible for the underlying behavioral momentum.
First Seen: when the system first detected this signal.
Because signals are behavior-based, they do not expire on a fixed schedule. Instead, they remain listed here until the underlying momentum dissipates, that is, when a token’s activity falls below the two-cluster threshold.
Recent alerts
The Recent Alerts section displays trade signals that were active within the past two hours but are no longer meeting the criteria for ongoing momentum.
In other words, these are alerts that recently lost traction — tokens that were trending across multiple clusters but have since cooled off or consolidated.
The information shown here is identical to the Active Alerts table, with one key difference:
Instead of First Seen, you’ll see Last Seen, which indicates how long ago the alert transitioned from active to inactive.
All other columns (Chain, Type, Strength, Price Target, Token, and Cluster list) remain the same, giving you consistent context for interpreting shifts in market behavior.
Recent Alerts help you track momentum fading in real time. Tokens that appear here may still be worth watching.
Prediction data display for active/recent alerts
Every active or recent alert presents a detail page with three separate categories of information.
Notification details
Price predictions
Trend start time information
This information, like all other aspects of the alerts will update periodically throughly the lifespan of a an alert. It will always show the must current information based on the latest network data.
Notification details
This section provides a complete snapshot of the alert at the moment it was generated, helping you understand why it appeared and which traders are driving the signal. Each label represents part of the alert’s underlying logic.
It contains basic information such as the name, ticker, type of alert, the clusters in which the alert is currently active and when it began.
Re-entry count measures how many times an alert has dropped out of active status and then reactivated within a short grace period. This happens when the token’s trending criteria are momentarily not met, for example, if cluster activity briefly dips below the threshold, but quickly recover.
Price Predictions
The prediction table shows nine independent forecasts for a single alert in a 3×3 grid of time horizons by price-change thresholds. It’s designed so you can see, at a glance, how far and how fast our models believe the move could go.
Rows will always capture the same 3 time periods (1 hour, 4 hours and 24 hours) and the amount of time left since the alert originated.
Remember, all price predictions are rolling, so a given time period may say Expired while still predicting a future price change. This countdown is presented to help emphasize the early period of an alert's lifespan, which is often associated with the greatest opportunities.
Columns capture 3 future price changes, where buy alerts will show price increases and sell alerts will show price decreases. These price levels will vary between high and low liquidity tokens.
Each cell shows the model’s probability (percentage + bar) and a banded label (e.g., High, Medium, Low Confidence) for a given price change and time-horizon combination. It will also contain a symbol representing a predicted price increase 
, price decrease 
or that price change not predicted to be met 
.
Here are several examples of a single cell from the predictions table and the correct interpretation for each one:
Here is one example of the complete prediction table from a buy alert and the correct interpretation of each cell from left-to-right then top-to-bottom in the image caption:
Here is one example of the complete prediction table from a sell alert and the correct interpretation of each cell from left-to-right then top-to-bottom in the image caption:
It should be clear from the above examples that the key elements of each prediction are:
The row label indicating the time horizon associated with the prediction, which is 1 hour
, 4 hours 
and 24 hours 
.The column label indicating the magnitude of price increase or decrease associated with the prediction, for example +10%
or -10% 
The price increase
, price decrease or no threshold not predicted to be met symbols associated with the prediction.The strength or confidence of the prediction, which ranges from 1% to 99% and is represented by a percentage, filled bar and banded label, for example
Hōkū price predictions are highly granular, but with some practice you'll learn to:
Scan the row that matches your trading horizon (1h, 4h, or 24h).
Glance across columns to see how aggressive the forecast is (5% → 30%).
Act on consistent strength (e.g., Medium/High across multiple thresholds).
We're constantly evolving the interface for displaying price predictions and value feedback from our users. Don't hesitate to reach out to the team if you have input on this module.
Trend start time
This section shows the start time of the alert in UTC, ISO and your local time as set by your browser.
Alert history
The alert history page shows a list of expired alerts that are more than 2 hours old. It is launched by clicking the "History" button at the top of the page.
While the Active Alerts list shows what’s happening right now, this section lets you look back at how well the models performed once those predictions played out.
Each entry in the list preserves the same core data you see in the Active view. The key addition is the Backtesting Result column, which is a metric unique to historical alerts.
Backtesting result
This value shows how many of the model’s 9 forecasts (the 3×3 grid of time horizons and price levels) were correct after the prediction window closed.
For example, a result of 7 / 9 
means 7 of the 9 predicted price movements occurred within the specified timeframes. Similarly, a result of 9/9 
means all 9 predictions were found to be correct.
These results are calculated automatically from live blockchain data, using real market prices at the exact prediction horizons. More details are shown in the prediction data display when you click a historical alert, which is discussed in the next section.
Prediction data display for historical alerts
Just like active and recent alerts, each historical alert show's 3 categories of information
Notification details
Validated price predictions
Trend start time information
The Notification details and Trend start time information essentially the same for historical alerts as active or recent alerts (for explanation on these categories see Prediction data display for active/recent alerts above). However, historical alerts show Validated price prediction information differs slightly because it clearly displays whether each prediction was correct or why it was not.
Validated price predictions
The Validated Price Prediction table builds upon the Price Predictions table described above, by providing details on the accuracy of the predictions on live market data. It is therefore only available on the alert history page and only once each of the 3 time horizons has elapsed. Like the Price Prediction table, it's designed so you can quickly see how accurate Hōkū's price predictions were for a given alert.
There are 3 key categories of information to take note of:
An overall summary that displays the number of correct predictions out of total predictions for which actual price data was available along with the number of predictions that are still pending
A narrative explanation that further describes the reasoning for any pending and/or incorrect predictions
The full 9-cell table that uses symbols to clearly display the status of each prediction. The actual price change for each time frame is shown on the left, and each cell contains a set of icons. For buy alerts, the green upward arrow
denotes a price increase, the red downward arrow 
denotes a price decrease, and the non-event symbol 
denotes the threshold not being met. Each timeframe and price threshold contains a set of 2 symbols (1 for the predicted price 1 for the actual observed price) , which when match indicate a correct prediction.
Keep in mind that a prediction can be marked corrected before its time frame has expired. If, for example, a 20% price increase is predicted for the 1 hour, 4 hour and 24 hour timeframes and an actual price increase of 20% occurs in the first hour, then all 3 time frames can be marked correct. However, the opposite scenario cannot occur. If a price threshold has not yet been met, the time frame must fully expire to definitively validate the prediction.
Below are 3 example tables along with corresponding explanations noted from left-to-right, top-to-bottom in the image captions. In each explanation, the green check ✅ indicates a correct prediction, the red X ❌ indicates an incorrect prediction, and clock 🕐 indicates a pending prediction whose timeframe has not expired.
This table shows a buy alert prediction that was highly accurate, with only 1 prediction pending.
This table shows a sell alert prediction that was accurate on the longer (4 hour and 24 hour) timeframes, but inaccurate on the shortest timeframe.
This table shows a buy alert prediction that was generally inaccurate but with the entire 24 hour timeframe still pending.
With some practice, you'll learn to quickly scan these tables for pairs of matching symbols, which indicate a correct prediction. But, just like the Price Prediction table data, you should always take into account the strength of each prediction, which is characterized by the confidence data for each prediction. You'll generally find that high confidence predictions are correct more often than low confidence predictions.
Usage tips
The price predictions associated with Hōkū’s alerts are clear, powerful trade signals, so the most obvious way to use them is just to trade! That said, we also came up with a few other strategies to consider.
Trade directly from prediction signals
The simplest way to use Trade Signal Alerts is to treat them as actionable trading cues. When Hōkū issues a Buy alert with a high confidence score, it means our behavioral models predict strong upward momentum — driven by real trader activity across multiple clusters. In practice, that often precedes price movement.
Likewise, a Sell alert indicates declining conviction and reduced buying pressure among key wallet groups. With the rise in availability of derivatives in crypto, even these are potential profit opportunities.
By aligning your trades with these signals you can capture market shifts as they begin, not after they’re obvious.
Validate your watchlist using backtested results
Use the Notification History tab to see how accurate recent predictions were for tokens you care about. If you notice that certain assets or clusters consistently achieve 7/9 or higher in backtesting, it means those behaviors are more predictable and may offer more reliable trading opportunities.
Over time, you can build a personalized list of “high-signal” tokens or clusters that historically align with your strategy.
Smarter capital rotation between short- and long-term positions
The time horizon filters in Hōkū make it easy to see when short-term or long-term opportunities dominate the market.
If the 1-hour prediction window is full of new, high-confidence signals, it often means the market is heating up — traders are moving fast, and capital is rotating rapidly between tokens. In these conditions, keeping too much capital in long-term positions can create opportunity cost.
Conversely, when most activity shifts toward 24-hour predictions, it signals a cooling market or the early stages of accumulation, creating conditions where holding stronger positions may make more sense. By monitoring the balance between short and long prediction activity, you can time when to stay agile and when to commit to conviction plays, using real on-chain behavior rather than sentiment or guesswork.
FAQ
How it works
Hōkū’s Trade Signal Alerts turn live cluster activity into rolling, forward-looking predictions. Unlike traditional TA (RSI, MACD, Bollinger Bands), these models forecast trader behavior, essentially who is likely to buy or sell, then directly estimate the probability of price reaching specific thresholds within 1h / 4h / 24h windows.
This pipeline runs separately on every supported network. For each chain, we routinely train nearly 4 million candidate models and only deploy the best performers. This training routine is refreshed periodically to counter model drift as market regimes evolve, which means in a given year, over 100 million individual machine learning models are trained to support Hōkū's unique price prediction features.
Step 1: Train the clustering models to assign addresses to clusters
We begin by classifying every active wallet into behavior-based clusters (which is explained further in the 3D Wallet Explorer docs). These clusters become the basic building blocks of our alert system training pipeline.
Step 2: Build labeled training examples from real alerts
We scan history for moments when a token met the alert criteria (e.g., trending across ≥2 clusters). For each such moment (“t₀”), we record the future outcomes: did price reach ±5% / ±10% / ±30% within 1h / 4h / 24h?
Each event becomes a binary label (reached / not reached, or event/non-event) for our training dataset. This is the variable our models learn to predict in the training process.
Step 3: Develop nearly 100 trader and market context features
For each training example (token at t₀), we compute a comprehensive feature set describing who is trading it and the environment they’re trading in. We can't say more about the specific features without revealing our secret sauce, so you'll have to use your imagination.
Step 4: Explore over 100,000+ model configurations for each model
Once all training data and features are ready, Hōkū runs a massive search to find the best way to translate behavioral signals into accurate predictions.
We use a boosted decision tree framework, which is a family of machine learning models that learns by combining thousands of tiny “if-then” rules into one powerful ensemble.
Each model configuration defines parameters such as how deep the trees can grow, how aggressively they learn from new examples, and how much randomness is injected to prevent overfitting. We generate a large grid of these settings amounting to over 100,000+ combinations per model then train them all in parallel.
Step 5: Train 36 specialized models for every prediction scenario
Each alert type (Buy or Sell) and each prediction target (1 h, 4 h, 24 h × small, medium, large move) is treated as its own task. This means Hōkū doesn’t rely on one monolithic model to predict everything and instead trains 36 independent specialists per network.
For each task, every one of the 100,000+ candidate configurations is evaluated. The best-performing configuration is selected based on accuracy, consistency, and speed in simulated real-time conditions.
While the selection process uses advanced statistical validation, in simple terms it’s like auditioning thousands of experts and keeping only those who consistently perform well under pressure.
Step 6: Deploy and run predictions in real time
Once the top 36 models are selected, they are loaded into Hōkū’s live inference system.
Every time a token meets the alert criteria (meaning it’s trending across at least two trader clusters) the system instantly builds an updated feature set and runs all relevant models to produce the 9-cell prediction grid in each alert.
Each prediction cell expresses the probability that the token will reach its target within its time window. The results are streamed instantly to the app through secure WebSocket connections, ensuring users see new signals the moment they’re computed.
Meanwhile, a parallel system continuously monitors how each prediction performs once time passes, automatically verifying accuracy and flagging model drift (when behavior patterns change). Models that lose calibration are retrained and replaced.
Last updated