Fraud Detection in Cryptocurrency Markets with Spatio-Temporal Graph Neural Networks Lidia Losavio∗ 1 , Luca Persia∗† 1 , Madan Sathe‡ , and Dimosthenis Pasadakis∗ ∗ Faculty of Informatics, Università della Svizzera italiana, Lugano, Switzerland † School of Management and Law, Zurich University of Applied Sciences, Zurich, Switzerland ‡ Forensics and Financial Crime, Deloitte AG, Switzerland
arXiv:2604.24590v1 [cs.LG] 27 Apr 2026
{lidia.anna.maria.losavio, luca.persia, dimosthenis.pasadakis}@usi.ch, [email protected]
Empirically, these ”campaigns” may leave distinct footprints in the price and the traded volume of the tokens. At the event level one typically observes: (i) a sudden increase in trading volume and number of trades over a short time interval, (ii) heavy concentration of activity on a single exchange or venue, and (iii) potential clustering of events in time, as the same groups may run multiple campaigns in quick succession. The existence of these distinct patterns in cryptocurrency market manipulation motivates a statistical learning approach for their early detection. The current state-of-the-art methods for detecting financial market frauds exploit the underlying structure of financial interactions [7]. They leverage graphical representations of financial data and deep learning, particularly graph neural networks (GNNs) [8], to capture relational patterns in the network [9], [10]. For cryptocurrencies in particular, fraud detection is increasingly graph-oriented, learning directly from interaction networks rather than just price or volume anomalies. In [11]–[13] transaction-level cryptocurrency information is exploited (e.g., sender-receiver directed graphs), which inherently implies a direct notion of the graphical structure of the problem. However, many accessible real-world datasets provide only aggregated market observations at the token level, such as OHLCV information (open, high, low, close prices and trading volume) over fixed time intervals, number of trades, and derived indicators. In this context, the relational structure between activities is not directly observed, as there are no explicit edges or direct notions of relational interactions. Consequently, a graph-based approach requires constructing an adjacency structure from aggregated time series, with different constructions leading to different inductive biases in the downstream GNN models.
Abstract—Technological advancements in cryptocurrency markets have increased accessibility for investors, but concurrently exposed them to the risks of market manipulations. Existing fraud detection mechanisms typically rely on machine learning methods that treat each financial asset (i.e., token) and its related transactions independently. However, market manipulation strategies are rarely isolated events, but are rather characterized by coordination, repetition, and frequent transfers among related assets. This suggests that relational structure constitutes an integral component of the signal and can be effectively represented through graphical means. In this paper, we propose three graph construction methods that rely on aggregated hourly market data. The proposed graphs are processed by a unified spatio-temporal Graph Neural Network (GNN) architecture that combines attention-based spatial aggregation with temporal Transformer encoding. We evaluate our methodology on a real-world dataset comprised of pump-and-dump schemes in cryptocurrency markets, spanning a period of over three years. Our comparative results showcase that our graph-based models achieve significant improvements over standard machine learning baselines in detecting anomalous events. Our work highlights that learned market connectivity provides substantial gains for detecting coordinated market manipulation schemes. Index Terms—spatio-temporal Graph Neural Network, market manipulation, pump-and-dump, cryptocurrency, fraud detection
I. I NTRODUCTION AND R ELATED W ORK Financial market manipulation is the attempt to distort the process by which the prices of financial assets are formed, so that they no longer reflect real information about supply and demand [1]. This encompasses many behaviors, but a useful research taxonomy distinguishes between trading-based, information-based and action-based manipulation [2]. These phenomena can be traced back to the stock market itself [2], yet there has recently been a notable surge in manipulative behaviour in the cryptocurrency markets. The fragmented liquidity and less stringent regulation of these markets may amplify the impact of manipulative activity, as discussed in [3]. Notable recent cases [4], [5] highlight that such fraudulent activities have a detrimental effect on both investors and on the credibility of the blockchain as an innovative system. The coordination of market manipulation in crypto markets often occurs on encrypted messaging platforms, where channels are established with the objective of attracting speculators [6].
Contributions and outline We introduce a framework that enables Spatio-Temporal GNN (ST-GNN) modeling through specific graph construction methods, utilizing only aggregated market data. To strictly evaluate and compare these construction strategies, we employ a unified ST-GNN predictor across all experiments. This architecture processes data in two stages: first, for each timestamp in a lookback window, a Transformer-style spatial GNN computes node embeddings based on the constructed
1 Equal contribution.
1
graph. Second, these spatial embeddings are stacked into a sequence, enriched with positional encodings, and processed by a temporal Transformer encoder to map the spatiotemporal context to a final event probability. For the graph construction, we follow a correlation-based construction approach [14] that quantifies co-movement in market aggregates during the training period. We adapt this idea to the problem of fraud detection in cryptocurrency markets, and derive the adjacency matrix from correlation coefficients, which we then sparsify via set thresholds. This graph construction approach is extended beyond a single static topology with two dynamic variants. Additionally, we consider a self-adaptive graph construction variant that learns latent dependencies directly from the data, capturing interactions beyond feature-based correlations. Our approach is evaluated on a real-world dataset of pump-and-dump (P&D) events extracted from [15] detected in the period 2017-2021. The three ST-GNN variants are compared against the widely used decision-tree methods Random Forests [16] and XGBoost [17]. Unlike existing P&D detection datasets, where the learning problem is formed by extracting a local window around each pump event and concatenating these labeled segments [6], [15], we consider a global historical window that is shared across all tokens in the sample. While local windows are well-suited to measuring event signatures, they also condition the sample on being near a pump, thereby discarding most daily market variation. This can reduce background market movements and may lead to an optimistic assessment of detection difficulty, i.e., a form of selection bias in the evaluation that results in high accuracy metrics. Our global window approach shifts the task toward a more realistic surveillance setting, i.e., a highly imbalanced learning problem in which fraudulent episodes are rare relative to the full stream of normal market activity. In particular, in our constructed dataset, we observe ∼ 0.016% of anomalous events in a total of ∼ 1.9 × 106 observations. The remainder of this paper is organized as follows. In Section II, we briefly recap the problem of detecting fraudulent events when assuming aggregated market data samples, and present a high level overview of our approach. In Section III we focus on our method for estimating graphs from these samples, and using them as input in a ST-GNN architecture to detect anomalous events. In Section IV, we perform numerical experiments on real-world data and compare with state-of-theart decision-tree methods in order to validate our proposed routines, and finally in Section V we draw conclusions from this work. Our code and data are publicly available at https: //github.com/lidialosavio-dotcom/crypto fraud stGNN.
II. F RAUD D ETECTION WITH AGGREGATED M ARKET DATA We address the detection of cryptocurrency fraud events at the token level, with hourly resolution using aggregated market data. In II-A we formally introduce the problem, and in II-B its formulation within a Graph Neural Network (GNN) architecture. A. Problem setting Let V = {1, . . . , N } be the set of traded tokens and let t ∈ {1, . . . , T } be the index of discrete hourly timestamps. For each token i ∈ V at time t, we observe a vector of aggregate features xi,t ∈ RF and a binary label yi,t ∈ {0, 1} indicating whether a fraudulent episode occurs for token i at time t. Let Xt ∈ RN ×F be the matrix containing all the feature vectors of the tokens at time t. Our goal is to learn a predictor fθ that estimates the probability p̂i,t of fraud for a specific token i at time t, p̂i,t = fθ (xi,t , xi,t−1 , . . . , xi,t−W +1 ) ∈ [0, 1], ŷi,t = I{p̂i,t ≥ γ},
(1)
where W is the lookback window and γ ∈ (0, 1) is a decision threshold to determine the binary prediction ŷi,t . In practice, xi,t typically includes OHLCV variables, trade counts, and additional engineered features, so that time-series information is available in tabular form. B. Formulation within a GNN architecture Graph Neural Networks (GNNs) are architectures designed to learn from graph-structured data by iteratively updating node representations through information exchange along edges (i.e., (k) message passing) [18]. Let hi ∈ RD denote the embedding (0) of node i after the k-th propagation step, with hi typically (0) initialized from the input features, i.e., hi = xi,t . A generic update then reads (k) (k−1) (k−1) hi = U(k) hi , AGG(k) {hj : j ∈ N(i)} , (2) where N(i) is the neighborhood of i, AGG(k) (·) is a permutation-invariant aggregating function over neighbors (e.g., sum/mean/max), and U(k) (·) is typically implemented by a Multilayer Perceptron (MLP). The number of steps k (k) controls the propagation depth; after k steps, hi incorporates information from nodes up to k hops away. In our setting, nodes correspond to traded tokens and edges encode inter-token relations derived from aggregated market dynamics. A spatio-temporal GNN (ST-GNN) extends message passing to time-indexed observations by combining a spatial graph aggregation at each time t with a temporal model that captures how node representations evolve across a lookback window [19]–[21]. Many GNN variants can be interpreted within this message-passing template, differing mainly in the way neighbor information is computed and propagated [22]– [24]. A practical constraint in cryptocurrency fraud detection arises from relying on exchange- or aggregator-provided APIs [25], [26] for data collection. These sources typically expose only
Notation In what follows, we denote scalar quantities with lowercase, vectors with lowercase bold, and matrices with uppercase bold characters. The (i, j)th entry of a matrix A is symbolized by Aij and all entries in row i or column j by Ai: and A:j , respectively. Sets are denoted by capital calligraphic characters, for example, A, the identity matrix as I and the vector of all ones as e.
2
market-level aggregates per token and time (e.g., OHLCV Algorithm 1 Graph inference strategies from aggregated data and related features) rather than transaction-level interactions. Input: Training timestamps Ttr , fraud timestamps Ptr , scalar Consequently, the relational structure between assets remains feature si,t , density ρ, min threshold τmin , dynamic horizon latent, as there are no explicit edges and any notion of L=12, self-adaptive threshold ϵ connectivity must be inferred from the observed time series [27]. Output: Inferred adjacency (static A, dynamic updates This limits the applicability of many graph-based approaches {A(p) }, or self-adaptive A) that rely on a pre-defined interaction network, thus motivating (G1) Static the development of methods capable of constructing graphs 1: Compute C on Ttr from log(1 + si,t ) from aggregated data, and utilizing them in the prediction task. 2: τ ← max(τmin , Qρ ({Cij }i<j )) 3: Aij ← Cij · I{Cij > τ } III. I NFERENCE G RAPHS M ETHODS AND A U NIFIED (G2) Dynamic event-driven S PATIO -T EMPORAL GNN 4: for each p ∈ Ptr in chronological order do We infer the graphical structure of token interactions from ag5: W(p) ← {p − L, . . . , p} gregated data, and utilize it in a spatio-temporal GNN to detect 6: Compute C(p) on W(p) (p) fraudulent events. Subsection III-A defines data representation 7: τ (p) ← max(τmin , Qρ ({Cij }i<j )) (p) and temporal windowing, and in Subsection III-B introduces 8: Add/update edges (i, j) where Cij > τ (p) using our graph inference strategies. Then, Subsection III-C details running mean the proposed GNN architecture. 9: Save A(p) 10: end for A. Data representation and temporal windowing (G3) Self-adaptive Given the node feature matrix Xt ∈ RN ×F , temporal context 11: Learn E1 , E2 ∈ RN ×d ; compute M ← E1 E⊤ 2 is captured through a fixed lookback window of length W . For 12: A ← softmax(ReLU(M)) (row-wise) each node i ∈ V at time t, we form the sequence 13: Sparsify: keep (i, j) only if Aij > ϵ (W ) W ×F Xi,t = xi,t−W +1 , . . . , xi,t ∈ R . (3) Stacking these sequences for all nodes yields the windowed (W ) (W ) tensor Xt ∈ RN ×W ×F , whose i-th slice satisfies Xt [i, : (W ) , :] = Xi,t . The model outputs a probability per node for the current time t, i.e., p̂t ∈ [0, 1]N .
volume or number of trades), we define the log-transformed token-by-time matrix S ∈ R|Ttr |×N with entries: St,i = log(1 + si,t ),
t ∈ Ttr , i ∈ {1, . . . , N }.
(5) N ×N
We compute the Pearson correlation matrix C ∈ [−1, 1] and set the diagonal entries to zero: Cij = Corr S:,i , S:,j , Cii = 0. (6)
B. Graph inference strategies We denote by At ∈ RN ×N a weighted adjacency matrix at time t, and by Et = {(i, j) : At,ij ̸= 0} the corresponding directed edge set. Undirected connections are treated as directed edges from i to j and vice versa. When required for message passing, we equivalently represent At as an edge list Et with associated edge weights at = {At,ij : (i, j) ∈ Et }. Since the underlying market graph is latent, we define an inference operator At = gt (D≤t ), (4)
To sparsify this matrix, we select a data-driven threshold τ via an off-diagonal quantile, subject to a minimum floor τmin : n o τ = max τmin , Qρ {Cij }i<j , Aij = Cij I{Cij > τ } , (7) with Aii = 0 and (i ̸= j). Here Qρ denotes the ρ-quantile of the off-diagonal entries. This threshold retains only the strongest edges, i.e., the top 1 − ρ fraction of them, while τmin which constructs At relying solely on information available up ensured that retained edges represent statistically meaningfull to time t. This ensures that the graph construction respects the correlations. temporal order of observations, and in practice, is restricted only Since the resulting adjacency matrix is time-invariant, we to training timestamps when building training-time graphs [28]. have At = A, ∀ t. This approach assumes that even without Here, D≤t denotes the observed aggregated market data up to explicit interactions, tokens can exhibit synchronized market time t, i.e., the time series of token-level data used to build Xt activity. A static correlation graph is the simplest relational and any scalar time series si,t employed by the graph inference inductive bias compatible with aggregated token features [29]. rule. We evaluate three graph construction methods for gt , and (G2) Event-driven dynamic correlation graph summarize them in Algorithm 1. Static graphs often fail to capture correlation shifts associated (G1) Static correlation graph with manipulation episodes, as they rely on a fixed represenWe construct a single weighted adjacency from the training tation of the data, and do not update over time [30]. Our period and reuse it at every timestamp. Let Ttr ⊆ T denote the second graph construction approach addresses this limitation set of training timestamps. Given a scalar time series si,t for by constructing a piecewise-constant adjacency matrix that each token i, obtained from a single chosen node feature (e.g., updates only around fraud events within the training data.
3
Let Ptr ⊆ Ttr be the set of training timestamps where at least one fraud event occurs. For each p ∈ Ptr , we consider a lookback window of L hours: W(p) = {p − L, . . . , p} ∩ Ttr .
To sparsify this dense graph, we set edges with weights below a small tuned constant ϵ to zero: E = {(i, j) : Aij > ϵ},
(8)
(p)
(13)
Unlike the undirected graphs in (G1) and (G2), this approach generates an unsymmetric matrix A ̸= A⊤ .
Note that L denotes a graph construction lookback window, distinct from the feature window W in Section III-A. Using the same scalar series si,t as in (G1), derived from volume or number of trades, we compute a window-specific correlation matrix C(p) from {log(1 + si,t ) : t ∈ W(p)}, enforcing (p) again Cii = 0. We then select edges representing significant interactions with a density-controlled threshold: n o (p) τ (p) = max τmin , Qρ {Cij }i<j , (9) E(p) = {(i, j) : i < j, Cij > τ (p) }.
Aij ← Aij I{Aij > ϵ},
C. ST-GNN architecture All inferred graphs are used within the same spatio-temporal model architecture. At each hour t, the inferred adjacency At defines the neighborhood Nt (i) = {j : At,ij ̸= 0}. We encode (i) spatial interactions on At with a Transformer-style attention GNN layer [32], and (ii) temporal dependencies via a Transformer encoder over an W -hour lookback window [20], [33]. First, spatial aggregation is performed at each time step. For each node i ∈ V, the input feature vector xi,t is mapped to a spatial embedding hi,t ∈ RD via attentive message passing over At :
(10)
For each undirected edge (i, j) we keep a cumulative sum Sij of (p) the correlation coefficients Cij , and a counter nij representing the number of fraud event windows where the edge (i, j) was part of the graph, i.e., (i, j) ∈ E(p) . The resulting edge weight hi,t = GraphTransformer(xi,t , At ). (14) is defined as the mean of these correlations, wij = Sij /nij . (p) The updated adjacency matrix A(p) is formed by setting Aij = Subsequently, we construct the temporal sequence by stacking (p) (p) the embeddings from the lookback window W : Aji = wij and Aij = 0 otherwise. At any time t, the model uses the graph state in the training Hi,t = hi,t−W +1 , . . . , hi,t ∈ RW ×D . (15) ⋆ set of the most recent fraud timestamp p , ( We add learnable positional encodings into Hi,t , and process I, if {p ∈ Ptr : p ≤ t} = ∅, the sequence with a temporal Transformer encoder, denoted At = (11) ⋆ A(p ) , p⋆ = max{p ∈ Ptr : p ≤ t}. by TE(·). The fraud event probability for node i at time t is For timestamps in the validation and testing sets, the graph obtained by A(plast ) of the last training update is considered, with plast = p̂i,t = σ w⊤ TE(Hi,t )[W ] + b , (16) max Ptr being the last fraud event timestamp in the training D where w ∈ R and b ∈ R are learnable output parameters and set. σ(·) is the sigmoid logistic function. Algorithm 2 summarizes (G3) Self-adaptive adjacency from node embeddings the forward pass. While correlation-based graphs capture statistical synchroIV. N UMERICAL R ESULTS nization between assets, they rely on fixed heuristic thresholds In this section, we present experimental results on real-world to determine connectivity, and thus, may not align with the optimal latent graph for fraud detection. To allow our data for an extensive evaluation of the accuracy and efficacy of model to discover latent dependencies directly from the data, the proposed graph-based fraud detection routines outlined in we employ a self-adaptive adjacency construction approach Section III. We will compare our methods against the following decision-tree based ensemble methods: inspired by [31]. We randomly initialize two learnable node embedding 1) Random Forests (RF) [16]: Aggregates predictions from dictionaries E1 , E2 ∈ RN ×d , representing source and target decorrelated decision trees trained on bootstrap samples node embeddings, respectively. These embeddings are optiand feature subsampling. We use a grid search to optimize mized together with the neural network weights via stochastic the number of estimators nest ∈ {500, 1000}), maximum gradient descent, guided solely by minimizing the supervised tree depth dmax ∈ {12, ∞}, minimum samples per leaf classificaiton objective. The hyperparameter d controls the rank nmin ∈ {1, 10}, and√the feature subset size considered at of the learned adjacency matrix A. The adjacency matrix, that each split Fsub ∈ { F , 0.8F }. can be considered as the transition matrix of a hidden diffusion 2) eXtreme Gradient Boosting (XGBoost) [17]: A gradientprocess [31], is then computed as boosted tree method that sequentially fits trees to correct previous errors. We tune the learning rate η ∈ {0.05, 0.1}, ⊤ A = SoftMax ReLU(E1 E2 ) . (12) maximum tree depth (dmax ∈ {4, 10}), number of esHere, the ReLU activation function eliminates weak or negative timators (nest ∈ {500, 1000}), and L2 regularization connections, while the row-wise SoftMax normalizes the parameter λ ∈ {1.0, 10.0}, while fixing row and column weights to obtain a row-stochastic adjacency. subsampling rates at 0.8.
4
BRD
OAX
03 -2 5
03 -2 6
Algorithm 2 ST-GNN forward pass (W )
Number of Trades
Input: Windowed node features Xt ∈ RN ×W ×F , inferred graph (Et , at ) where at are edge weights, parameters θ Output: Event probabilities p̂t ∈ [0, 1]N 1: seq ← [ ] ▷ will store spatial embeddings for each hour in the window 2: for u = 1 to W do ▷ iterate over window steps (W ) 3: H ← Win Xt [:, u, :] ▷ linear projection 4: H ← TransformerConv1 (H, Et , at ) 5: H ← ReLU(H); H ← Dropout(H) 6: H ← TransformerConv2 (H, Et , at ) 7: H ← ReLU(H) 8: seq.append(H) 9: end for 10: S ← Stack(seq) ∈ RN ×W ×D 11: S ← S + P ▷ positional embedding 12: S̃ ← TemporalTransformerEncoder(S) 13: z ← S̃[:, W, :] ▷ last time token 14: p̂t ← σ(Wout z + b) 15: return p̂t
6,000 4,000 2,000
03 -2 7
03 -2 4
03 -2 3
0
Time Fig. 1: Hourly number of trades for BRD and OAX over a four-day window around one pump event.
to the set of pump timestamps associated with that token. The dataset contains 1,905,850 observations for N = 84 tokens. The positive class (flagged pump hours) includes 314 positives, The following Subsection IV-A summarizes our experimental corresponding to a positive rate of ∼ 0.016%. setup, and then IV-B is devoted in a comparison of the accuracy We compute 18 additional features on this base panel to of the methods under consideration. capture abrupt regime shifts, typical of coordinated manipulation. Since our dataset is limited to candle-level aggregates A. Experimental setup and does not include order-book information (e.g., bid–ask We consider a real-world dataset of pump-and-dump events depth or spread), we construct a proxy for aggressive buy from, covering the period July 2017 – January 2021, where pressure using the ratio between taker buy quote volume and labels where assigned by monitoring encrypted messaging total traded volume at the candle level. For each token, we platforms and linking each market manipulation campaign to compute rolling statistics over a fixed window of 12 hours and a target token [34]. We collect the available market data for then consider their percentage change over time, following the quote currency BTC through the Binance API [25], and assign feature construction in [34]. Such percentage changes of rolling a unique timestamp by combining the date and hour fields and means or rolling standard deviations quantify how quickly the localizing the resulting time to UTC. Since market data are local average or local variability is shifting, which is the type represented as hourly candles1 , pump timestamps are rounded of signal expected under coordinated activity. Following this to the nearest hour boundary to enforce a consistent alignment approach, we compute percentage changes of rolling mean between the pump schedule and the candle grid. A pump and standard deviation for the buy pressure, traded volume, episode from our dataset is illustrated in Figure 1, where the number of trades, and close price. We then compute the number of trades spikes sharply around the event time for two same transformation for the rolling mean of the intraperiod tokens, highlighting both the market signature of coordinated maximum price. Last, since intraday patterns can be relevant activity and the presence of cross-token co-movement that when campaign timing is systematic, we derive an hour feature motivates our graph-based approach. as the integer hour extracted from the timestamp. The full list For each token we download a continuous time interval of the features considered is offered in Table I in Section VI spanning from the earliest to the latest pump timestamp in the of the Appendix. schedule, both extended by a fixed margin of seven days. To respect temporal ordering and emulate a surveillance Consequently, all tokens share the same time boundaries, setting, we adopt a chronological hold-out split for all considresulting in a panel that includes both normal market activity ered models. Let T denote the ordered set of unique hourly and rare fraud episodes, contains the candle timestamp, OHLCV timestamps in the panel. We partition T into 60% for training, information and trading activity. The binary fraud label flag is 20% for validation, and 20% for testing. Graph construction assigned by checking whether the candle timestamp belongs using historical observations is performed using only the training data, model hyperparameters and early stopping criteria 1 In financial market microstructures, a candle (or candlestick) aggregates are selected on the validation set. The decision threshold γ is trading activity over a fixed time interval and is typically represented by the also selected on the validation set by maximizing the F1-score, OHLC information, together with volume and related trade counts for that interval. and the same threshold is applied unchanged to the test set,
5
XGBoost
RF
Static (Volume)
Dyn (Volume)
Dyn (Num. Trades)
Self-Adaptive
0.6
0.9
0.7
0.7
0.6
0.6
0.5 Precision
0.5 F1 score
Static (Num. Trades)
0.4 0.3
0.4 0.3 0.2
0.2
0.1
0.1
0.5 Method (a)
0.7 0.8 Recall
1
(b)
Fig. 2: Classification performance for all methods under consideration. (a) F1-score, and (b) mean precision-recall curves for Recall ∈ [0.5, 1].
which is utilized only for the final accuracy estimation. Then, to mitigate short-range temporal dependence and reduce leakage around split boundaries, an embargo of z = 5 hours is adopted and the first z timestamps of the validation and test blocks are discarded [35]. We evaluate performance in terms of precision, recall, and F1 score, defined as: F1-score = 2 ·
Precision · Recall , Precision + Recall
methods achieve the lowest scores, with XGBoost (in black) yielding the lowest F1 = 0.49 ± 0.03 and Random Forests (in purple) improving upon it with F1 = 0.53 ± 0.01. All our introduced graph-based variants outperform these baselines in terms of classification accuracy. The static graph based on volume correlation (in brown), and its dynamic counterpart (in orange) achieve F1 = 0.575 ± 0.04 and F1 = 0.58 ± 0.02, respectively. The strongest correlation-based performance is obtained when connectivity is derived from the number of trades as a proxy for trading activity with F1 = 0.58 ± 0.04 for the dynamic graph (in blue) and F1 = 0.60 ± 0.04) for the static one (in green). The self-adaptive graph (in red) achieves the best overall accuracy with F 1 = 0.62 ± 0.05, indicating that allowing the model to adjust connectivity during training is beneficial when the graph topology is not available. The moderate error bars across seeds confirm that these rankings are stable and not driven by favorable initializations. To further validate that the observed gains stem from the graph structure rather than from increased model capacity, we evaluate temporal-only baselines (GRU, TransformerEncoder) that process each token independently; full details are provided in Appendix VI-C. A comparison of model sizes and training times is reported in Table III in Appendix VI-D.
(17)
where precision is defined as Precision = TP/(TP + FP) and recall as Recall = TP/(TP + FN), with TP, FP, and FN denoting true positives, false positives, and false negatives classifications, respectively. All the results reported hereby correspond to their mean value after 9 runs. For all experiments, we employ the ST-GNN architecture defined in Section III-C with a fixed learning rate of 10−3 , and a spatial embedding dimension of D = 64 (Eq. 14), 2 attention heads, and a temporal window of W = 5 hours. Regarding graph construction, for the correlation-based strategies (G1, G2), we set the minimum threshold to τmin = 0.15, the dynamic lookback window to L = 12. The quantile density in (G1) is set to ρ = 0.90 when considering the number of trades, to 0.75 when considering the volume, and to 0.95 for both dynamic variants in (G1). Finally, for the self adaptive graph (G3), the learnable embedding dimension is d = 48 and the sparsity threshold is ϵ = 0.005. The complete grid search spaces and selected configurations for all graph inference strategies are reported in Table II in Appendix VI-B. All the results reported hereby correspond to their mean value after 9 runs.
In Figure 2b, we report the mean Precision-Recall (PR) curves for Recall ∈ [0.5, 1]. Given the extreme class imbalance in our test set, containing only ∼ 0.016% positive pump hours in a total of ∼ 1.9 × 106 hours, we prioritize PR curves over ROC analysis, as the latter can be insensitive to the false-positive burden when negatives dominate [36]. The plot highlights a consistent performance gap, with graph-based models shifting the PR frontier upward relative to the tree-based baselines, implying higher precision for any given recall level.
B. Classification performance We present in Figure 2a the F1-score (17) of the classification of the minority (i.e., fraudulent) class. The two tree-based
6
XGBoost
RF
Static (Volume)
Dyn (Volume)
Dyn (Num. Trades)
Static (Num. Trades)
Self-Adaptive
1
F1 score
0.8 0.6 0.4 0.2
A
N
X
S
PP C
0
Fig. 3: Classification performance in terms of F1-score for all methods under consideration for tokens (i.e., APPC and NXS) with at least five pump events in the test set.
While performance naturally degrades as recall approaches 1, algorithms, and their applicability for the monitoring of market our graph-based methods maintain precision levels significantly manipulation events. longer than the tree-based baselines. The best method is the self-adaptive variant, which dominates the frontier across the ACKNOWLEDGMENT majority of the sensitivity range. We acknowledge the financial supported by the joint DFGIn Figure 3 we present the classification performance in terms of F1-score (17) for the most affected tokens, with at 470857344 and SNSF-204817 project entitled “Numerical Alleast five recorded pump events in the test set. All graph- gorithms, Frameworks, and Scalable Technologies for Extremebased variants introduced in Subection III-B report accuracy Scale Computing”, and by the Huawei Zurich Research scores higher than the competing tree-based methods. The Center. This work was also funded by the “Resilient Financial self-adaptive variant (in red) achieves the highest scores with Enterprises” project, supported by the Swiss Innovation Agency F1 = 0.808 ± 0.05 for APPC and F1 = 0.931 ± 0.06 for NXS, (Innosuisse) under grant agreement 119.321 INT-ICT. where APPC and NXS are individual token ticker symbols, each representing a single node in the graph. R EFERENCES [1] Netherlands Authority for the Financial Markets (AFM), “Market manipulation,” Regulatory brochure (Market Abuse Regulation context), 2017, brochure processed ESMA Q&As up to 6 July 2017. [2] T. J. Putniņš, “Market manipulation: A survey,” Journal of Economic Surveys, vol. 26, no. 5, pp. 952–967, 2012. [3] J. T. Hamrick, F. Rouhi, A. Mukherjee, A. Feder, N. Gandal, T. Moore, and M. Vasek, “An examination of the cryptocurrency pump-and-dump ecosystem,” Information Processing and Management, vol. 58, no. 4, p. 102610, 2021. [4] S. Shifflett and P. Vigna, “Some traders are talking up cryptocurrencies, then dumping them, costing others millions,” The Wall Street Journal, Aug. 2018, available at: https://www.wsj.com/graphics/ cryptocurrency-schemes-generate-big-coin/. [5] B. Merewhuader, C. Polite, and T. Viane, “Class action complaint, merewhuader et al. v. safemoon llc et al., no. 2:22-cv-01108 (c.d. cal. feb. 17, 2022),” Filed in U.S. District Court for the Central District of California, 2022. [6] J. Xu and B. Livshits, “The anatomy of a cryptocurrency Pump-and-Dump scheme,” in 28th USENIX Security Symposium (USENIX Security 19). Santa Clara, CA: USENIX Association, 2019, pp. 1609–1625. [7] L. Akoglu, H. Tong, and D. Koutra, “Graph based anomaly detection and description: a survey,” Data Mining and Knowledge Discovery, vol. 29, no. 3, pp. 626–688, 2015. [8] X. Ma, J. Wu, S. Xue, J. Yang, C. Zhou, Q. Z. Sheng, H. Xiong, and L. Akoglu, “A comprehensive survey on graph anomaly detection with deep learning,” IEEE transactions on knowledge and data engineering, vol. 35, no. 12, pp. 12 012–12 038, 2021.
V. C ONCLUSIONS We presented a framework to detect manipulation in cryptocurrency markets, specifically addressing the scenario where only aggregated market data are available, i.e., OHLCV and trading activity, while transaction-level information remains unobservable. To overcome this limitation, we proposed a methodology that infers token-to-token connectivity from market dynamics, and processes it via a unified spatio-temporal GNN (ST-GNN) architecture. We proposed three graph inference methods: static correlation, event-driven dynamic correlation, and self-adaptive adjacency learning. We compared our graph-based solutions against tree-based methods on a real-world labeled dataset of pump-and-dump schemes spanning the period 2017–2021. Our results demonstrate that graph-based models are superior in terms of F1-score and Precision-Recall efficiency. Notably, the selfadaptive adjacency method achieved the highest fraud detection performance, confirming that the model can successfully learn latent interaction structure between tokens. Our work highlights the effectiveness of the introduced graph-based fraud detection
7
[9] Y. Dou, Z. Liu, L. Sun, Y. Deng, H. Peng, and P. S. Yu, “Enhancing graph neural network-based fraud detectors against camouflaged fraudsters,” in Proceedings of the 29th ACM International Conference on Information & Knowledge Management (CIKM ’20). Virtual Event, Ireland: Association for Computing Machinery, Oct. 2020. [10] J. Schmidt, D. Pasadakis, M. Sathe, and O. Schenk, “GAMLNet: a graph based framework for the detection of money laundering,” in 2024 11th IEEE Swiss Conference on Data Science (SDS), 2024, pp. 241–245. [11] C. Wu, J. Chen, Z. Zhao, K. He, G. Xu, Y. Wu, H. Wang, H. Li, Y. Liu, and Y. Xiang, “Tokenscout: Early detection of ethereum scam tokens via temporal graph learning,” in Proceedings of the 2024 on ACM SIGSAC Conference on Computer and Communications Security, CCS 2024, Salt Lake City, UT, USA, October 14-18, 2024, B. Luo, X. Liao, J. Xu, E. Kirda, and D. Lie, Eds. ACM, 2024, pp. 956–970. [12] C. Wu, J. Chen, J. Li, J. Xu, J. Jia, Y. Hu, Y. Feng, Y. Liu, and Y. Xiang, “Profit or deceit? mitigating pump and dump in defi via graph and contrastive learning,” IEEE Transactions on Information Forensics and Security, vol. 20, pp. 8994–9008, 2025. [13] Z. Lin, Q. Luo, D. Wu, J. Shen, L. Li, X. Nong, and Z. Qin, “Detecting illicit transactions in bitcoin: a wavelet-temporal graph transformer approach for anti-money laundering,” Scientific Reports, vol. 16, no. 1, p. 1548, 2026. [14] K. da Costa, “Anomaly detection in global financial markets with graph neural networks and nonextensive entropy,” 2023, arXiv:2308.02914v2. [15] M. La Morgia, A. Mei, F. Sassi, and J. Stefa, “The pump and dump scheme in the bitcoin era: Real time detection of cryptocurrency market manipulations,” in 2020 29th International Conference on Computer Communications and Networks (ICCCN). IEEE, 2020, pp. 1–9. [16] L. Breiman, “Random forests,” Machine Learning, vol. 45, no. 1, pp. 5–32, 2001. [17] T. Chen and C. Guestrin, “Xgboost: A scalable tree boosting system,” in Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, ser. KDD ’16. New York, NY, USA: Association for Computing Machinery, 2016, p. 785–794. [18] G. Corso, H. Stark, S. Jegelka, T. Jaakkola, and R. Barzilay, “Graph neural networks,” Nature Reviews Methods Primers, vol. 4, no. 1, p. 17, 2024. [19] B. Yu, H. Yin, and Z. Zhu, “Spatio-temporal graph convolutional networks: A deep learning framework for traffic forecasting,” in Proceedings of the 27th International Joint Conference on Artificial Intelligence, 2018, pp. 3634–3640. [20] G. Jin, Y. Liang, Y. Fang, Z. Shao, J. Huang, J. Zhang, and Y. Zheng, “Spatio-temporal graph neural networks for predictive learning in urban computing: A survey,” IEEE transactions on knowledge and data engineering, vol. 36, no. 10, pp. 5388–5408, 2023. [21] Z. Wu, S. Pan, G. Long, J. Jiang, X. Chang, and C. Zhang, “Connecting the dots: Multivariate time series forecasting with graph neural networks,” in Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, 2020, pp. 753–763. [22] T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” in 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings, 2017. [23] P. Velickovic, G. Cucurull, A. Casanova, A. Romero, P. Liò, and Y. Bengio, “Graph attention networks,” in 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings, 2018. [24] K. Xu, W. Hu, J. Leskovec, and S. Jegelka, “How powerful are graph neural networks?” in 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019, 2019. [25] Binance, “Binance spot api documentation: Rest api,” https://developers. binance.com/docs/binance-spot-api-docs/rest-api, accessed: 2026-02-01. [26] CoinGecko, “Coingecko api documentation,” https://docs.coingecko.com/ reference/introduction, accessed: 2026-02-01. [27] D. Pasadakis, M. Bollhöfer, and O. Schenk, “Sparse quadratic approximation for graph learning,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 9, pp. 11 256–11 269, 2023. [28] C. Shang, J. Chen, and J. Bi, “Discrete graph structure learning for forecasting multiple time series,” in International Conference on Learning Representations (ICLR), 2021. [29] M. Tumminello, T. Aste, T. Di Matteo, and R. N. Mantegna, “Correlation, hierarchies, and networks in financial markets,” Journal of Economic Behavior & Organization, vol. 75, no. 1, pp. 40–58, 2010.
[30] D. Matsunaga, T. Suzumura, and T. Takahashi, “Exploring graph neural networks for stock market predictions with rolling window analysis,” CoRR, vol. abs/1909.10660, 2019. [31] Z. Wu, S. Pan, G. Long, J. Jiang, and C. Zhang, “Graph wavenet for deep spatial-temporal graph modeling,” in Proceedings of the 28th International Joint Conference on Artificial Intelligence, ser. IJCAI’19. AAAI Press, 2019, p. 1907–1913. [32] Y. Shi, Z. Huang, S. Feng, H. Zhong, W. Wang, and Y. Sun, “Masked label prediction: Unified message passing model for semi-supervised classification,” in Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21, Z.-H. Zhou, Ed. International Joint Conferences on Artificial Intelligence Organization, 8 2021, pp. 1548– 1554, main Track. [33] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” Advances in Neural Information Processing Systems, vol. 30, 2017. [34] M. La Morgia, A. Mei, F. Sassi, and J. Stefa, “Pump and dumps in the bitcoin era: Real time detection of cryptocurrency market manipulations,” in 2020 29th International Conference on Computer Communications and Networks (ICCCN), 2020, pp. 1–9. [35] M. López de Prado, Advances in Financial Machine Learning. Hoboken, NJ: John Wiley & Sons, 2018. [36] T. Saito and M. Rehmsmeier, “The precision-recall plot is more informative than the roc plot when evaluating binary classifiers on imbalanced datasets,” PloS one, vol. 10, no. 3, p. e0118432, 2015.
8
VI. A PPENDIX
Signal
Search Space
Selected
dropout ∈ {0.2, 0.3, 0.4} ρ ∈ {0.95, 0.90, 0.75} dropout ∈ {0.2, 0.3, 0.4} ρ ∈ {0.95, 0.90, 0.75}
dropout = 0.3 ρ = 0.90 dropout = 0.2 ρ = 0.75
dropout ∈ {0.2, 0.3, 0.4} L ∈ {12, 24} ρ ∈ {0.95, 0.90, 0.75} dropout ∈ {0.2, 0.3, 0.4} L ∈ {12, 24} ρ ∈ {0.95, 0.90, 0.75}
dropout = 0.3 L = 12 ρ = 0.95 dropout = 0.2 L = 12 ρ = 0.95
dropout ∈ {0.2, 0.3, 0.4} d ∈ {16, 32, 48, 64} ϵ ∈ {0.005, 0.001, 0.02}
dropout = 0.4 d = 48 ϵ = 0.005
(G1) Static
A. Features
si,t = num_trades
Feature
Type Definition
OHLCV quote asset volume num trades taker buy base taker buy quote std rush order avg rush order
raw raw raw raw raw eng eng
std trades std volume std price avg volume avg price avg price max hour of the day
eng eng eng eng eng eng time
si,t = volume
Open/high/low/close and volume. Volume in quote currency. Trades per hour. Hourly taker-buy volume (base). Hourly taker-buy volume (quote). Pct. change 12h roll std (buy pressure). Pct. change 12h roll mean (buy pressure). Pct. change 12h roll std (num trades). Pct. change 12h roll std (volume). Pct. change 12h roll std (close). Pct. change 12h roll mean (volume). Pct. change 12h roll mean (close). Pct. change 12h roll mean (high). UTC hour of the day.
(G2) Dynamic si,t = num_trades
si,t = volume
(G3) Self-Adaptive
TABLE II. Grid search spaces and final selected hyperparameters for the three graph inference strategies. For (G1) and (G2), the scalar series si,t used to construct the correlation-based graphs is reported.
TABLE I. Feature set used in the numerical experiment. Each token corresponds to a graph node i ∈ V, and the features listed here constitute the node feature vector xi,t ∈ RF at each hourly timestamp. Raw features are obtained directly from the market data downloader, engineered features (eng) are constructed during the feature engineering stage, and temporal features (time) are derived from the UTC timestamp of each observation.
D. Computational footprint and training time Table III reports the approximate model size and training time of the main baselines and the proposed graph-based variants. Overall, (G1) Static provides the best trade-off between performance and computational cost. Although (G3) Self-Adaptive is the most expensive among the proposed graph-based models, it remains practically deployable and is comparable to stronger non-neural baselines in wall-clock time. All experiments were executed on a single NVIDIA RTX A6000 GPU, with each model trained in isolation (full GPU availability).
B. Hyperparameter selection for graph inference To isolate the effect of the graph inference strategy, the ST-GNN architecture was kept fixed across all experiments, and only the graph-inference hyperparameters and the dropout probability were tuned. For all methods, the learning rate was fixed to 10−3 . For the correlation-based graphs (G1) and (G2), the scalar series si,t used to construct the graph was chosen from a single raw feature, namely num_trades or volume. The final selected configurations are reported in Table II. The ST-GNN with (G1) and (G2) contains 135,041 learnable parameters, whereas (G3) contains 143,105 due to the additional learnable node embedding matrices E1 and E2 . Parameter counts are also summarized in Table III. C. Temporal-only baselines To further assess the contribution of the inferred graph structure, we also evaluated two temporal-only baselines, namely a GRU and a TransformerEncoder, which process token histories independently and therefore cannot model crosstoken coordination. Among them, the GRU was the strongest temporal-only baseline, achieving a maximum F1-score of 0.56 and a PR-AUC of 0.51. This remains below the best graphbased model, for which the self-adaptive graph (G3) reaches a PR-AUC of 0.58. Importantly, increasing the GRU capacity from approximately 16k to 164k trainable parameters did not close the gap, indicating that the improvement of the proposed approach does not arise merely from higher model capacity, but from the use of an inferred relational structure that enables cross-token information exchange.
Model
Approx. Num. Parameters
Approx. Training Time
GRU (small) GRU (large) TransformerEncoder (G1) Static (G2) Dynamic (G3) Self-Adaptive XGBoost Random Forest
∼ 16k ∼ 164k ∼ 64k ∼ 135k ∼ 135k ∼ 143k ∼ 15k ∼ 500k
∼ 3 min ∼ 4 min ∼ 12 min ∼ 15 min ∼ 31 min ∼ 58 min ∼ 60 min ∼ 60 min
TABLE III. Approximate model size and training time of the temporal-only baselines, graph-based variants, and classical baselines. Times are wall-clock approximations under the same experimental setup. For classical machine learning baselines, we report approximate model size rather than neural trainable parameters.
9