Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/alphaleaks60-maker/docs2/llms.txt

Use this file to discover all available pages before exploring further.

Phase 1 establishes the ground truth about every wallet the system observes. It answers a single question: given everything this wallet has ever done on Pump.fun, how good are they? The alpha scores, realized PnL figures, and feature vectors produced here are the most important inputs to the ML models in Phase 3, and they determine which signals get promoted to the live trader.

FeatureComputer

30-minute rolling feature profiles covering entry behaviour, sizing, holds, and returns.

WalletScorer

7-component alpha score (0–100) derived from the computed feature set.

PeakTracker

Retrospective 1h, 4h, and 24h price peak measurements per signal.

PnlCalculator

FIFO-matched realized PnL, win rate, and hold time per wallet per token.

FeatureComputer

The FeatureComputer runs every 30 minutes and processes up to 5,000 wallets that have been active in the last 35 minutes, or that have never had features computed and have bought at least 3 tokens. Features are computed with point-in-time correctness — they reflect only what was known at the moment of observation, not retroactively adjusted data. For each eligible wallet, the computed feature set is written into wallet_features and covers five behavioural dimensions:
These features measure how quickly and how early a wallet enters positions after token launch.
FeatureDescription
avg_seconds_after_creationAverage time in seconds between token launch and this wallet’s first buy
pct_buys_under_60sShare of buys placed within the first 60 seconds of a token’s life
pct_buys_under_300sShare of buys placed within the first 300 seconds of a token’s life
pct_top10_entriesHow often this wallet is among the first 10 buyers on a token
Early entry is strongly correlated with better outcomes on Pump.fun, because the bonding curve price is lowest at launch. Wallets that consistently enter in the first 60 seconds — and in the top 10 buyers — are identified as potential alpha sources.
These features characterise how a wallet sizes its positions.
FeatureDescription
avg_sol_per_buyMean SOL spent per buy transaction
median_sol_per_buyMedian SOL spent per buy (more robust to outliers)
sol_size_stddevStandard deviation of buy sizes — measures consistency
max_single_buyLargest single buy ever recorded for this wallet
High variance in position sizing can indicate opportunistic scaling into conviction trades. Low variance suggests a systematic, rules-driven approach.
These features describe how long a wallet holds positions before selling.
FeatureDescription
avg_hold_time_secondsAverage time between the first buy and the final sell for a position
pct_quick_flipsShare of positions fully exited within 5 minutes of opening
pct_diamond_handsShare of positions held longer than 24 hours
Quick-flip wallets are often momentum scalpers; longer holders tend to target graduation events. Both archetypes can be alpha sources, but the live trader’s strategy configuration may prefer one over the other.
These features capture the breadth of the wallet’s token selection.
FeatureDescription
unique_creators_tradedTotal number of distinct creators whose tokens this wallet has bought
pct_repeat_creator_buysShare of buys directed at creators this wallet has traded before
A high pct_repeat_creator_buys can indicate an insider relationship with specific creators — useful context for the CreatorRiskScorer in Phase 2.
Activity features measure recency and breadth; return features are sourced from PnlCalculator and PeakTracker.Activity
FeatureDescription
active_days_30dNumber of distinct calendar days active in the last 30 days
tokens_tradedTotal unique tokens bought across all time
Returns
FeatureDescription
graduation_rateShare of tokens this wallet bought that eventually graduated to Raydium
win_rateShare of closed positions that were profitable
sol_weighted_returnReturn weighted by position size in SOL
avg_realized_multipleAverage realized return multiple across all closed positions
avg_peak_multipleAverage peak price multiple reached by tokens this wallet bought
capture_efficiencyRatio of realized return to peak available return (how close to the top the wallet sold)
profit_factorGross profit divided by gross loss
return_stddevStandard deviation of per-trade returns
avg_loss_pct_on_losersAverage percentage loss on positions that closed at a loss
avg_gain_pct_on_winnersAverage percentage gain on positions that closed at a profit

WalletScorer

The WalletScorer runs every 30 minutes on wallets whose features have been updated since their last scoring. A wallet must have traded at least 10 tokens to produce a valid score — below that threshold, the sample size is too small for reliable ranking. The output is an alpha score between 0 and 100, computed from seven independent components. Each component is normalised and then weighted before being summed.
ComponentWeightDescription
Graduation rate20%Share of tokens bought that eventually graduated to Raydium. Normalised against a 50% ceiling — a 50% graduation rate maps to 100% on this component.
Win rate20%Share of positions closed in profit. Straightforward fraction of winning to total closed positions.
Capture efficiency15%How much of the available peak return the wallet actually realised. A wallet that consistently sells near the top scores highly.
Return consistency15%Inverse of return standard deviation. High variance penalises the score — a consistently profitable wallet outranks a lottery-ticket player with the same average.
Buy rank10%How early the wallet typically enters relative to other buyers on the same token. Earlier average entry yields a higher component score.
Recency10%Exponential decay applied to wallets that have been inactive. Wallets inactive for 14 or more days begin losing score on this component.
Discovery10%Average peak price multiple of tokens bought by this wallet. Rewards wallets that identify tokens before the crowd drives the price up.
Wallets classified as bots by the BotDetector service are capped at a maximum alpha score of 30, regardless of how well they score on individual components. The confidence value (0–1) scales with the number of tokens traded, reaching full confidence at 50 tokens.

How the alpha score feeds Phase 3

The alpha score is included directly in the 68-feature ML vector passed to MlInference. It is also the primary input to the rule-based SignalScorer in Phase 2. A wallet with an alpha score below the configured threshold will not trigger a signal at all — the score gates entry into the signal pipeline before any ML inference runs.

PeakTracker

The PeakTracker retrospectively measures how high a token’s price went after a signal was emitted. It runs continuously and updates signals with peak measurements at three time horizons:
FieldDescription
peak_multiple_1hHighest price multiple reached within 1 hour of the signal
peak_multiple_4hHighest price multiple reached within 4 hours of the signal
peak_multiple_24hHighest price multiple reached within 24 hours of the signal
In addition to the continuous multiples, the PeakTracker writes binary target labels — for example, reached_2x_1h, reached_3x_30m — which are the ground truth labels used to train the ML models in Phase 3.
Peak measurements are the bridge between real-time signal generation and offline ML training. Without accurate peak labels, the models have no ground truth to learn from. The quality of labels directly determines the quality of the trained models.

PnlCalculator

The PnlCalculator pairs every buy event with its corresponding sell events on a per-wallet, per-token basis to compute realised profit and loss. It uses a FIFO (first in, first out) matching approach: the earliest open position is closed first when a sell is observed. For each matched position, it records:
  • Realised multiple (exit price / entry price)
  • Hold time in seconds from first buy to final sell
  • Whether the position closed at a profit or a loss
These per-position records feed directly into FeatureComputer (for aggregated wallet statistics) and WalletScorer (for win rate, return consistency, and profit factor components).

OutcomeTracker

The OutcomeTracker closes the feedback loop for signal quality. Every signal emitted by the pipeline is checked against actual on-chain outcomes at 1-hour, 4-hour, and 24-hour intervals after emission. This outcome data feeds three consumers:
  • PeakTracker — to write the peak multiple and binary label fields
  • ML training pipeline — as the labelled dataset for model training
  • ModelMonitor in Phase 3 — to detect drift between model predictions and real outcomes

Phase 2: Signal intelligence

See how Phase 1 wallet scores feed into token lifecycle, bundle detection, and creator risk scoring.

ML features

Full reference for the 68-feature vector built from Phase 1 and Phase 2 outputs.

Build docs developers (and LLMs) love