FinML-Chain: A Blockchain-Integrated Dataset for Enhanced Financial Machine Learning

Data

Collection for On-chain Data

We collect the data through BigQuery, and the code we used is in Query

You can also refer to BigQuery for more information.

Collection for Off-chain Data

On-chain Data Infomation

Data Files	Data Type	Data Content
ETH-Token-airdrop.csv	Raw Data	Critical indicators related to gas during tokrn airdrop period
ETH-Normal.csv	Raw Data	Critical indicators related to gas during normal period

On chain Data Dictionary

ETH-Token-airdrop.csv and ETH-Normal.csv

Variable Name	Description	Type
timestamp	Recoding of the time of each block	String
number	The number of blocks on the chain	Numeric
gas_used	Actual gas used	Numeric
gas_limit	The maximum allowed gas per block	Numeric
base_fee_per_gas	The base fee set for each block	Numeric

Additional Variables we create

Variable Name	Description	Type
gas_fraction	Fraction between Gas Used and Gas Limit	Numeric
gas_target	The optimal gas used for each block	Numeric
Y	Normalized Gas Used	Numeric
Y_t	Response variable equals to the gas_fraction	Numeric

Off-chain Data Information

Variable Name	Description	Type
chat text	people's chat (sentences)	String

Code

Code Files	Code Description
main_dataset_processing_code.ipynb	Applying FinBert to process discord information; Applying the NAM model to manipulate monotonicity; Applying Both on-chain data and off-chain data to train the model
NAM models.py	NAM model
baseline_dataset_processing_code.ipynb	Using linear algorithm, DNN, XGBoost and long-short term memory to predict gas used.

Results

Baseline results

Baseline loss for Token-airdrop period		Baseline loss for Token-airdrop period
Baseline variance for Token-airdrop period		Baseline variance for Token-airdrop period

Baseline loss for normal period		Baseline loss for normal period
Baseline variance for normal period		Baseline variance for normal period

Flow chart

Flow chart of combination of Off-chain and On-chain

Flow chart of combination of Off-chain and On-chain

Monotonicity Two-step training loss (normal training and monotonic training)

We utilized the NAM model due to its inherent transparency characteristic and the ability to isolate variables, facilitating the imposition of monotonicity constraints on specific features. The model is trained on data from two distinct periods, achieving weak pairwise monotonicity over the $\alpha$ feature. In the first step, standard training is conducted to enable the model to learn from the data. In the second step, we impose monotonic constraints.

Two-step training loss

Sentiment (Combination of Off-chain and On-chain)

We further explore the NAM model at k=1,2 and 3. Given the availability of both on-chain and off-chain variables, we conducted tests to determine whether the inclusion of off-chain variables, specifically sentiment analysis, enhances the model's predictability.

Model Performance over Two Periods

Model Performance over Two Periods
	+OC,+DS,+HS	+OC,+DS,-HS	+OC,-DS,+HS	+OC,-DS,-HS
Period 1: 03/21/2023 - 04/01/2023 (ARB-airdrop)
3 Timesteps	0.10022	0.10150	0.10164	0.10201
2 Timesteps	0.10056	0.10249	0.10213	0.10265
1 Timestep	0.10169	0.10190	0.10204	0.10290
Period 2: 06/01/2023 - 07/01/2023 (Normal)
3 Timesteps	0.13341	0.15657	0.16142	0.16089
2 Timesteps	0.13477	0.15381	0.15806	0.16456
1 Timestep	0.13593	0.15321	0.15459	0.18428

The notation "OC" refers to On-chain variables, while "HS" and "DS" denote Hourly Averaged Sentiment and Daily Averaged Sentiment, respectively. The ‘+’ symbol indicates the inclusion of a variable in the model, whereas the ‘-’ symbol denotes its exclusion. The numerical values represent the mean square error (MSE) of the model on the test dataset.