Background

Over the last couple of months, we have been studying how gas fees change over time for few of the prominent blockchain networks. This analysis was expected to provide insights into what may help decide for the best time for crypto trading over a day, a week, or a random seasonal period.

We were hoping our effort will also give us clues towards predicting gas fees for the immediate 24-48 hours that would help in deciding on some key points of a trade’s lifecycle.

So, what are gas fees?

Gas fees are a charge incurred for using a blockchain network like Ethereum.

When one wants to make a transaction on a blockchain, such that the transaction incurs gas fees, one would need to pay gas fees to the network alongside the transaction. Essentially, these fees are paid to the miners on the network who run the network’s computation thereby maintaining the network secure and stable.

The amount of gas fees one need to pay depends on the complexity of the transaction and the current demand for network resources. Think of it like paying a toll to use a highway - the more traffic there is, the higher the toll.

As a practitioner in any industry using blockchain, understanding gas fees in detail is essential - the knowledge helps to estimate the cost of using blockchain networks and planning project design with measures to control fees.

For example, a crypto investor could save gas fees in transactions by designing a strategy that uses blockchains with lower fees. Or the investor uses gas fees information to prioritise specific transactions at chosen times, favourable for their strategy.

In another example, a tokenisation platform that makes transactions such as burning or minting of tokens, or sending tokens between wallets, could optimise on achieving those transactions by appropriately addressing gas fees requirements either through the blockchain selection process, or within the smart contract specification.

Our Approach

For the sake of experimental research, we chose to assess the most popular of blockchains out there - Bitcoin and Ethereum. Similar to stable-coins, we expected relatively less volatility in their blockchain fees. (Bitcoin’s fees aren’t exactly called ‘gas fees’ - but conceptually it is similar to that of Ethereum). The biggest hurdle here was to find manageable amounts of historical gas data that was reliable and inexpensive to use.

Sourcing historical data from the blockchain

After looking at a couple of options for sourcing blockchain data, we ended up using Dune (dune.com) as our primary data source. Dune’s SQL queries make it easier to aggregate data, and its free subscription allows for a certain number of exports that were just enough for this analysis.

Dune’s dataset is also one of the most reliable and comprehensive datasets available, with both, API and SQL access mechanisms.

Example: SQL query to extract hourly average gas price data for BTC from Dune’s table (bitcoin.transactions)

We started with collecting gas fees data for an entire year, aggregating it over an hourly scale. The data reflected the min, max and mean of gas fees along with the number of rows (transactions) that occurred in that hour.

With Ethereum, the data set being so extensive that it contains both successful and unsuccessful transactions, we had to filter out the unsuccessful transactions, thus giving us with a cleaner dataset.

Initial observations

To get an initial understanding of the gas fees data set, we plotted time-series graphs for average fees and count of transactions for the 1 year for both BTC and ETH side by side (click to enlarge).

We can see that the average fees for Bitcoin had a huge spike in May 2023 that overshadowed other smaller spikes. As against this time range, Ethereum has been experiencing gas fee spikes on a regular basis but none have been as exorbitant as that of BTC, at least in the last year dataset.

Note that the units for fees is not the same for both BTC and ETH, hence both graphs aren’t directly comparable.

With the number of transactions, it is apparent that Ethereum is a busier network than Bitcoin. That being said, Bitcoin’s transactions have taken a jump again around May 2023.

We found that the correlations between average gas fees, min gas fees, max gas fees, in Ethereum’s dataset indicated that the minimum gas price for an hour is very closely correlated to the average gas price, implying that the prices seldom soar from the average, but when they do they can be very high as reflected in the max gas price dataset.

In the case of Ethereum, we observe that there was little influence of the network congestion to the average gas prices, the number of transactions did not correlate with any aspect of gas fees.

Diving deeper into the 24-hour trend

Next, we analysed how the gas prices for Ethereum behaved over 24 hours cumulatively. We plotted the mean prices for each hour averaging over the entire year and found some interesting results.

It was observable that on average, between 04:00 UTC - 11:00 UTC, gas prices remain fairly low and then start rising until reaching their peak at 16:00 UTC before declining till the end of the day. This was an important insight as it can be useful for placing on chain transactions strategically on particular hours to minimize the amount of gas fees one pays for a transaction.

Similar results were obtained when we plotted each hour’s average gas prices as a scatter plot for the 24 hours. Disregarding some outliers, the 24 hour trend above seems fairly consistent.

We then plotted an auto-correlation function (ACF) plot to find autocorrelation for gas prices over a lagged timeline.

As seen in the plot, lags 1 to 5 strongly auto-correlate in the average gas prices. This means that the current gas price depends closely on the gas prices a few hours ago. However, we can notice that while there is a decline for further lags, the autocorrelation once again reaches the peak at 24 lags. This implies strong correlation between current gas prices and the gas price exactly 24 hours before it. If followed further, we notice similar behaviour after every 24 hour time range.

The results further corroborate our findings above with the 24 hours trend.

Why does Ethereum gas fees peak at 16:00 UTC?

As puzzling as it may seem, there appears to be a possibility of certain markets or institutions forcing their on-chain transactions to complete at “any cost” before the midnight or end of day has elapsed in Asia region (UTC + 8 Hours) or in Europe (UTC + 1). This may be an area for further exploration.

Gas Fees Prediction Modelling

We created an ARIMA model using the Autocorrelation as well as Moving Averages mechanism to predict gas prices over the next few days. We limited our training dataset to end as of 1st July and we attempted to predict the gas fees for the immediate 24 - 72 hours.

While the initial predictions were able to capture the highs and lows of the trend, they weren’t close enough to the actual gas prices for July 2023. And with further data points, the trend changed to a straight line unable to predict the behaviour of the gas prices. We believe with more complex prediction modelling techniques, these shortcomings could be improved upon.

Conclusion

There seems to be no lack of historical data in crypto, especially when it comes to the major public blockchains. Disregarding the occasional and seasonal macro highs and lows, there are some solid trends that can be observed throughout these datasets using basic statistical techniques.

With the crypto analytics ecosystem expanding every day, it is only a matter of time that we can access more fundamental information of block-chain activity as well as of price trend in near real-time.

Investors and tokenisers may be able to use such information to predict token prices and gas fees over a narrow time window with greater confidence and exercise control on when and how on-chain transactions are executed.

Bibliography

• [ACF Plot] - Interpreting ACF and PACF Plots for Time Series Forecasting 

• Using Python and Auto ARIMA to Forecast Seasonal Time Series

• Stationarity Assumption in Time Series Data

• Weekly, daily and sub-daily data | Forecasting

• API - Dune Docs