Languages: English | 繁體中文 | 简体中文

Are Ethereum Whales Smart Money? An Event Study of On-Chain Signals and Sentiment

Abstract

This study investigates whether large on-chain Ethereum transactions predict short-term and medium-term ETH price movements, and whether market sentiment moderates this relationship. Using an event study methodology on 646,442 whale transactions over 3.5 years (Jan 2023 - Jul 2026), we measure the directional hit rate of whale exchange deposits (sell signals) and withdrawals (buy signals) across nine time horizons (1 hour to 6 months), four transaction size thresholds ($1M to $10M+), seven sentiment regimes, two market regimes (bull/bear), and four independent yearly samples.

Live App: https://crypto-whale-signals-and-sentiment-lkhygb3594bbrogn23qbps.streamlit.app/

Key findings:

The deposit (sell) signal's edge grows monotonically from +1.3% at 24 hours to +12.4% at 6 months, and this pattern was set using 2023-2025 data and confirmed, unchanged, when tested against 2026 data collected afterward.

1. Whale deposits (sell signals) show a persistent, growing edge that strengthens with horizon, and holds at every transaction size. The unconditional deposit edge grows from +1.3% at 24h to +4.8% at 1 month to +12.4% at 6 months. This growth over time shows up consistently whether we look at $1M+ or $10M+ deposits alone: both move from roughly flat in 2023 to +2.9% to +3.9% by 2026. Whale sellers think in weeks and months, not hours. Explicitly conditioning the signal on sentiment (extreme greed) only helps at 24h, lifting the edge by +1.3% to +2.2% depending on transaction size; at every longer horizon tested, that same conditioning weakens the edge instead, by as much as 17.3% at 6 months, consistently across all four thresholds. The unconditional signal, not the sentiment-filtered one, is what holds up over time. (A separate, narrower claim we initially reported, that $10M+ deposits during extreme greed showed a 78.3% hit rate, did not hold up under closer testing; see Discussion.)

2. Whale withdrawals (buy signals) lost their edge entirely, and this holds at every transaction size too. In 2023-2024, withdrawals during negative funding showed +4.7% to +10.1% edge; by 2025 that had faded close to zero at every threshold, and by 2026 it was negative at every threshold from $1M+ to $10M+, ruling out the simple explanation that small, less sophisticated actors were just diluting the signal. At longer horizons, withdrawals become actively wrong: -6.1% edge at 1 month, -12.9% at 6 months. We attribute this to DeFi maturation: withdrawals increasingly represent staking, liquidity provision, and L2 bridging rather than directional buying.

3. Alpha decay is asymmetric, and a headwind that should have weakened the deposit signal did not. As ETH's price rose, a fixed $1M threshold captured progressively smaller, less committed sellers (~833 ETH in 2023 vs ~250 ETH in 2026), dilution that should work against the deposit edge. Instead the edge grew, which argues against a simple composition-effect explanation. One candidate explanation for why the deposit signal avoided being arbitraged away, unlike withdrawals, is that whale-watching tools amplify bullish activity more than bearish.

4. The edge is real, and it comes with a real cost. For long-horizon deposit signals that eventually paid off, we measured the maximum adverse excursion: the worst unrealised loss before the signal worked. This grows sharply with horizon, from 2.7% at 1 week to 20.1% at 6 months on average, with the worst 10% of "correct" 6-month trades seeing a 54.4% adverse move first. The edge is not free money; collecting it means surviving drawdowns most traders cannot tolerate.

---

Table of Contents

• Background
• Original Contributions
• Data Sources
• Methodology
• Results: Unconditional Hit Rates
• Results: Conditioned on Sentiment
• Results: Yearly Stability
• Results: Threshold Sensitivity
• Results: Long Horizons
• Results: Deposits by Year at Long Horizons
• Results: Drawdown During the Holding Period
• Results: Bull vs Bear Market Regimes
• Discussion
• Limitations
• Related Academic Literature
• Dashboard
• Repository Structure
• How to Run

---

Background

In traditional equity markets, institutional positioning is reported quarterly (13F filings) with significant lag. In contrast, public blockchains like Ethereum have every transaction visible in near-real-time. This creates a unique opportunity to study large-holder behaviours in real time.

Prior work on "whale watching" treated it as a price prediction problem: feed whale data into an ML model and attempt to forecast returns. This approach conflates multiple signals and rarely isolates the whale-specific contribution. This new approach now differs in three ways:

1. Event study methodology: we directly measure whether whale actions predict direction, rather than building a black-box predictor.
2. Sentiment conditioning: we test whether market regime (Fear & Greed Index, futures funding rate) moderates the whale signal.
3. Yearly stability analysis. We test each year independently (2023, 2024, 2025, 2026) to detect alpha decay. 2026 serves as an out-of-sample check on hypotheses developed from 2023 to 2025 data.

Original Contributions

1. Asymmetric alpha decay. Whale buy signals decayed from 2023 to 2026 while sell signals strengthened. We hypothesise this is partly because whale-watching tools broadcast bullish activity more than bearish, but the cause is not definitively established.

2. Long-horizon whale intelligence. Whale deposit edge grows monotonically with time horizon: +1.3% at 24h, +4.8% at 1 month, +12.4% at 6 months. The longer-horizon figures carry more overlap between events and deserve more caution than the 24h figure (see Limitations). Whale sellers are not just day-traders, they can have better experience and knowledge and might see structural shifts weeks to months ahead.

3. Threshold sensitivity, revised. We initially reported $10M+ deposits during extreme greed achieving a 78.3% hit rate at 24h. Looking closer, the transactions behind that number cluster within a small number of days, which is the same overlapping-observations problem described in the Limitations section, and it is why we do not treat this figure as confirmed (see Discussion).

4. DeFi dilution hypothesis. Due to DeFi maturation, withdrawals can mean staking, LP provision, and L2 bridging, which are non-directional. This dilutes the "withdrawal = buy" assumption. However, deposits remain a clean sell signal.

5. Out-of-sample test. All hypotheses were developed on 2023 to 2025 data. 2026 data was then fetched separately and tested without any parameter tuning. Deposit signals survive; withdrawal signals do not.

---

Data Sources

| Source | Data | Records | Period | Cost | |--------|------|---------|--------|------| | Dune Analytics | Whale transactions (>$1M) | 646,442 | Jan 2023 to Jul 2026 | Free tier | | Binance API | Hourly ETH/USDT prices | 30,801 | Jan 2023 to Jul 2026 | Free | | Binance API | ETH funding rates (8-hourly) | 3,851 | Jan 2023 to Jul 2026 | Free | | alternative.me | Crypto Fear & Greed Index | 3,075 | Feb 2018 to Jul 2026 | Free | | Kaggle | Bitcoin news headlines | 5,906 | Jan 2023 to Sep 2024 | Free | | GitHub (open-source) | Wallet address labels | 52,768 | Snapshot | Free |

Methodology

Data Pipeline (Phase 1)

• 646,442 large ETH transactions via Dune Analytics SQL query
• 52,768 wallet addresses labelled by merging brianleect/etherscan-labels and dawsbot/eth-labels
• Label coverage: 62.8% of transactions have at least one identified address
• MEV bot candidates flagged for sensitivity analysis

Transaction Classification (Phase 2)

• Rule-based labelling for known wallets (exchange deposit, withdrawal, DeFi, wallet-to-wallet)
• An ML classifier (Random Forest) predicts categories for unknown wallets, achieving 67.7% accuracy on a time-based hold-out (see scripts/run_phase2_classifier_eval.py). After expanding the label dataset from 30 to 52,768 addresses, label coverage reached 62.8%, reducing reliance on the classifier. It is still used for the remaining ~37% of transactions where both sender and receiver are unknown. This is the same task Harlev et al. (2018) tackle for Bitcoin entity types (10 categories, ~200M transactions): their best result was 77% accuracy with Gradient Boosting, modestly higher than this Random Forest's 67.7%, though the two differ in algorithm, features, and dataset

Category distribution:

| Category | Count | Share | Meaning | |----------|-------|-------|---------| | wallet_to_wallet | 321,257 | 49.7% | Both addresses are non-exchange, non-DeFi. Could be OTC trades, cold storage transfers, or personal wallet shuffling. No directional signal because intent is unknown, so these are excluded from the event study | | exchange_deposit | 181,105 | 28.0% | ETH sent TO a known exchange. Interpreted as a sell signal. The main reason to deposit to an exchange is to sell | | exchange_withdrawal | 124,772 | 19.3% | ETH withdrawn FROM a known exchange. Historically interpreted as a buy signal (accumulating), though this assumption has weakened as DeFi matured | | defi_interaction | 19,308 | 3.0% | One address is a kn