Feesim is a Bitcoin fee estimation program. Unlike fee estimation in Bitcoin Core, Feesim makes use of the current mempool state and transaction arrival rate, allowing it to be responsive to traffic surges. The fee estimates can be seen here.

Fee estimation in Feesim is model-based; estimates are obtained from Monte Carlo simulation of a queueing model of the bitcoin network. In essence, we find the solution to the question: "If we assume that miners prioritize transactions by fee rate, subject to a max block size and min fee rate, and that transactions arrive at the same rate as the past hour, then given the current mempool size, what fee rate is required for a transaction to be confirmed within X blocks (with success probability P)?"

This allows the estimation algorithm to accommodate variations in network conditions. For example, if a block hasn't been found in an hour and there's a large mempool backlog, the algorithm should recognize that and raise the fee estimates. Alternately, if there are daily lulls in transaction arrival rate, the fee estimates should reflect that as well, allowing the user to pay lower fees during the lull periods.

Feesim collects Bitcoin network data through the Bitcoin Core JSON-RPC API, and computes estimates for:

• The network hashrate
• The distribution of max block size and min fee rate* ("mining policy")
• Short-term transaction arrival rate / joint distribution of (txFeerate, txSize)

*See commit d895e64

Together with the current mempool state, these are used to perform the simulation and obtain the fee estimates. Mining policy estimates can be seen here, and transaction arrival byte rates as a function of fee rate can be seen here.

During normal operation, Feesim will predict the confirmation time of each transaction and keep a tally of the proportion of transactions which met the prediction. This proportion should be close to the success probability (default 90%), if the model is accurate. These scores can be be seen here.

Install from source using at least Go 1.6:

$ go get github.com/bitcoinfees/feesim

Feesim uses the Go 1.5 vendor experiment, so alternatively you can install with Go 1.5 by setting the environment variable GO15VENDOREXPERIMENT=1.

Feesim requires JSON-RPC access to a Bitcoin Core node (version >0.13.0, which can be pruned). The RPC settings should be specified in config.yml, as such:

bitcoinrpc:
    username: myusername
    password: mypassword
    # host: localhost
    # port: 8332

The config file should be placed in the data directory:

• Linux: ~/.feesim
• OS X: ~/Library/Application Support/Feesim
• Windows: %LOCALAPPDATA%\Feesim

Upon running feesim start, the program will start data collection, and then begin running the simulation once there is sufficient data. It needs to be online all the time, as it collects mempool data which cannot be obtained by offline analysis.

feesim status shows the program status:

$ feesim status
result      : Tx estimation window size was 0s, should be at least 600s
txsource    : Tx estimation window size was 0s, should be at least 600s
blocksource : Block coverage was only 0/2016, should be at least 1008/2016.
mempool     : OK

result shows whether or not fee estimates are available. By default, fee estimates require at least 10 minutes of transaction data, and data from 1008 of the last 2016 blocks.

Once there is sufficient data, the simulation will begin to run and produce fee estimates. The interface mirrors that of bitcoin-cli estimatefee:

$ feesim estimatefee 1
0.00030112

This shows the minimum fee rate for a transaction to be confirmed in 1 block, with 90% probability (configurable).

Unlike bitcoin-cli, if the input argument is ommitted or is 0, the estimates for all confirmation times is returned:

$ feesim estimatefee
 1: 0.00030138
 2: 0.00026738
 3: 0.00020492
 4: 0.00015988
 5: 0.00012805
 6: 0.00011478
 7: 0.00010616
 8: 0.00010001
 9: 0.00007519
10: 0.00005020
11: 0.00005000
12: 0.00005000

The underlying JSON-RPC API is compatible with Bitcoin Core's, so Feesim can be used as drop-in replacement for the estimatefee API:

$ bitcoin-cli -rpcport=8350 estimatefee 1
0.00030138

Feesim currently assumes that miners have the same minrelaytxfee as your node, so don't set it too low, or else wait time estimates for low fees will be inaccurate. Staying with the defaults should be OK. See commit d895e64 for details.

You may want to configure Bitcoin Core's maxmempool to be lower than the default, as this will make the simulation faster and require less memory. maxmempool=100 is fine if you stick with the default feesim settings (specifically maxblockconfirms). THe downside is having less comprehensive mempool data.

The simulation is CPU intensive, whereas data collection is not, so you may not want to run the sim all the time, while still collecting data. To do this, use feesim pause to pause the simulation, and feesim unpause to resume.

By default, fee estimates are updated every minute. It's possible, however, that a single simulation run takes longer than a minute, due to insufficient CPU resources or exceptionally high transaction traffic. In general, this will not cause any major problems; it only causes the fee estimates to be updated less regularly. It's possible, nevertheless, to reduce the simulation run time by lowering maxblocksconfirms or numiters in the config.

You can monitor the simulation run time with feesim metrics; sim.X are the run time statistics, in nanoseconds, for roughly the last X simulation runs.

Please see config.yml in this repository for an example config file.

As mentioned earlier, Feesim requires, by default, data from 1008 of the past 2016 blocks (it must be online when the blocks are discovered in order for the data to be logged). This is about 1 week; if you don't want to wait that long, you can contact me for a copy of the block data.