Better fee estimates #10199

morcos · 2017-04-12T17:18:26Z

This is a substantial improvement of fee estimation code built off #9942.
Although in many cases the estimates will not appear to differ by much from previous estimates, they will handle some situations much better.

A longer writeup of the original algorithm and the changes in this PR can be found here:
https://gist.github.com/morcos/d3637f015bc4e607e1fd10d8351e9f41

A summary of major changes:

Estimates can now be given for up to a 1008 block target.
Estimates are tracked on 3 different time horizons which allows for both the longer targets and lets the estimates adjust more quickly to changes in conditions.
estimatefee is deprecated in favor of using only estimatesmartfee (already used by the GUI)
estimateSmartFee now requires a 60% success rate at target/2, an 85% rate at target and a 95% rate at 2*target.
estimates are by default conservative which requires the 95% rate for 2*target to be met at longer time horizons as well, but by choosing non-conservative estimates, the estimates are drastically reduced during periods of less transaction activity (such as the weekend).
estimates are smarter about making sure enough data has been gathered in order to return a valid estimate (at least twice the number of blocks as the requested target much have been recorded)
Fee rate buckets are half as big leading to a bit more precision.
estimaterawfee is added so that customized logic can be implemented to analyze the raw data and calculate estimates.
Transactions which leave the mempool due to eviction or other non-confirmed reasons now count as failures. These are also saved at shutdown.

A summary of open issues:

Estimates are not backwards compatible. Currently when updating to this patch, any old fee estimates will be discarded and your node will have to be running for a while to provide new estimates. It is a possibility to add logic to return the old estimates for targets 2-25 until the new estimates have gathered enough data.
estimatefee is probably a roughly similar calculation to what it used to be, but I've made no effort to make sure the quality of the estimates it gives have not degraded, so infrastructure which doesn't support estimatesmartfee should update.
The GUI needs to be updated to access the new range of estimates. I would suggest providing options for targets of: 2,4,6,12,24,48,144,504,1008

jonasschnelli · 2017-04-13T11:58:45Z

Great!
Concept ACK.

The GUI needs to be updated to access the new range of estimates. I would suggest providing options for targets of: 2,4,6,12,24,48,144,504,1008

I'll have a look at this soon.

jonasschnelli · 2017-04-19T17:03:44Z

ping @crwatkins (he presented some fee estimation overview at the wallet standards group meeting in Berlin this month).

jlopp · 2017-04-19T19:22:04Z

src/rpc/mining.cpp

+            "\nWARNING: This interface is unstable and may disappear or change!\n"
+            "\nWARNING: This is an advanced API call that is tightly coupled to the specific\n"
+            "         implementation of fee estimation. The parameters it can be called with\n"
+            "         and the results it returns will change if the internal implementaion changes.\n"


Nit: implementation

jlopp · 2017-04-19T19:23:56Z

I'm not sure if this is a problem specific to your code, though requiring a leading 0 seems like a bug:

bitcoin-cli estimaterawfee 10 .85 0
error: Error parsing JSON:.85

sipa · 2017-04-19T19:49:56Z

@jlopp .85 is not valid JSON.

jlopp · 2017-04-19T20:18:07Z

src/rpc/mining.cpp

+    int nBlocks = request.params[0].get_int();
+    double threshold = 0.95;
+    if (!request.params[1].isNull())
+        threshold = request.params[1].get_real();


Would prefer more explicit range bounding on the threshold with an "Invalid threshold" error response; it currently allows numbers < 0 and > 1 and just returns a -1 feerate.

The reason I didn't bother with that is I felt estimaterawfee should be a very low level feature that doesn't need all the user friendly bells and whistles.. But it could certainly be cleaned up to add that either in this PR or a follow on..

Move buckets and bucketMap to be stored as part of overall serialization of estimator. Add some placeholder data so file format is only changed once. Maintain 3 different TxConfirmStats with potential for different decays and scales.

morcos · 2017-04-20T20:10:29Z

clean rebase after #9942 was merged

crwatkins · 2017-04-24T23:13:35Z

Thanks @jonasschnelli. In Berlin, I noted that at least 50% of the wallets listed on bitcoin.org use Core as a fee estimation source directly or through other servers based on Core and any statistical or algorithmic biases might compound over time because of the large number of clients using the single source. I presented some anecdotal, non-scientific observations about fees and some reasons that they might be artificially high. Some of my concerns about the current esitmatefee were

Slow to respond (2.5 day half life)
Too high confidence (95%)
No long term estimates (only 25 blocks)

I believe that the changes in this PR mitigate some of those concerns, and thus I am in favor of these changes. Nice work!

I'm still a little concerned about so many wallets "competing" using the same identical algorithm. @morcos do you envision some devs (wallets) would use estimaterawfee to tweak the algorithm for their users?

jlopp · 2017-04-25T00:33:07Z

@crwatkins I can only speak for BitGo, but we run custom compiled versions of Core with our own half life and confidence values. We then use the output as a baseline for a more complex fee algorithm that we adjust upward and downward based upon other data sources. I'm already working on transitioning us to using the estimaterawfee functionality.

morcos · 2017-04-25T17:05:14Z

@crwatkins Yes that was the basic idea of including estimaterawfee (that and it's useful for debugging). The exact strategy you use for determining what fee you want to put on a transaction depends on so many factors that its hard to have a one size fits all solution, but thats exactly what Bitcoin Core has to try to ship. So rather than attempt to defend why the algorithm I picked is the one algorithm to rule them all, I thought it would also make sense to expose a way to examine all the collected data and let people build their own algorithms if they'd like to.

crwatkins · 2017-04-25T17:08:36Z

@morcos Exclellent. Thanks!

Leviathn · 2017-04-25T18:05:39Z

Reviewing

TheBlueMatt

Looks good. With the exception of the fallback during upgrade and the one strange sharp edge LGTM. Needs testing because the code being correct isnt really enough, but no doubt better than today's estimates.

TheBlueMatt · 2017-04-25T18:51:19Z

src/policy/fees.cpp


    // Start counting from highest(default) or lowest feerate transactions
    for (int bucket = startbucket; bucket >= 0 && bucket <= maxbucketindex; bucket += step) {
+        if (newBucketRange) {


Can you update the commit message from "Track start of new bucket range more carefully" to something more descriptive.

TheBlueMatt · 2017-04-25T19:37:42Z

src/rpc/mining.cpp

+            "  \"leftmempool\" : x.x,    (numeric) number of txs over history horizon in the feerate range that left mempool unconfirmed after target\n"
+            "}\n"
+            "\n"
+            "A negative value is returned if no answer can be given.\n"


nit: "A negative feerate is returned" (other values are zero, so might as well just specify feerate here).

TheBlueMatt · 2017-04-25T22:58:51Z

src/policy/fees.cpp

-             100 * nConf / (totalNum + extraNum), nConf, totalNum, extraNum);
+    // If we were passing until we reached last few buckets with insufficient data, then report those as failed
+    if (passing && !newBucketRange) {
+        unsigned int failMinBucket = curNearBucket < curFarBucket ? curNearBucket : curFarBucket;


nit: std::min/max here like you did above?

TheBlueMatt · 2017-04-25T23:10:44Z

src/policy/fees.cpp

-    std::vector<std::vector<int> > curBlockConf; // curBlockConf[Y][X]
+    std::vector<std::vector<double>> confAvg; // confAvg[Y][X]
+
+    std::vector<std::vector<double>> failAvg; // future use


Should probably update this comment in the commit that puts failAvg to use.

TheBlueMatt · 2017-04-27T18:03:25Z

src/policy/fees.cpp

+            std::map<double, unsigned int> tempMap;
+
+            std::unique_ptr<TxConfirmStats> tempFeeStats(new TxConfirmStats(tempBuckets, tempMap, MAX_BLOCK_CONFIRMS, tempDecay));
+            tempFeeStats->Read(filein, nVersionThatWrote, tempNum);


As discussed it'd be nice to use this chunk and tempFeeStats to just generate one fee estimate, cache that, and then start fresh (as you do) but use the cached fee estimate while our new buckets fill.

leaving for a later improvement

TheBlueMatt · 2017-04-27T18:08:58Z

src/policy/fees.cpp

@@ -465,6 +469,8 @@ void CBlockPolicyEstimator::processTransaction(const CTxMemPoolEntry& entry, boo

    mapMemPoolTxs[hash].blockHeight = txHeight;
    mapMemPoolTxs[hash].bucketIndex = feeStats->NewTx(txHeight, (double)feeRate.GetFeePerK());
+    shortStats->NewTx(txHeight, (double)feeRate.GetFeePerK());


nit: might be nice to assert that the bucketIndex is the same across all three calls here?

TheBlueMatt · 2017-04-27T18:52:56Z

src/policy/fees.h


    /** Write estimation data to a file */
    bool Write(CAutoFile& fileout) const;

    /** Read estimation data from a file */
    bool Read(CAutoFile& filein);

+    /** Empty mempool transactions on shutdown to record failure to confirm for txs still in mempool */
+    void ClearInMempool(CTxMemPool& pool);


nit: maybe a better name for this would be FlushFailedPreShutdown()?

TheBlueMatt · 2017-04-27T20:32:01Z

src/txmempool.cpp

@@ -448,7 +448,7 @@ void CTxMemPool::removeUnchecked(txiter it, MemPoolRemovalReason reason)
    mapLinks.erase(it);
    mapTx.erase(it);
    nTransactionsUpdated++;
-    if (minerPolicyEstimator) {minerPolicyEstimator->removeTx(hash);}
+    if (minerPolicyEstimator) {minerPolicyEstimator->removeTx(hash, false);}


nit: it seems kinda strange that the "inBlock" flag is always set to false, even though it is sometimes called when the transaction was, indeed, in a block. Not a bug because it is a dummy call in fees.cpp, but maybe the API should stay the same and have the flag filled in inside of fees.cpp.

This will eventually be cleaned up when fee estimation listens to CValidationInterface

TheBlueMatt · 2017-04-27T23:25:06Z

src/policy/fees.cpp

+            return CFeeRate(0);
+
+        if ((unsigned int)confTarget > maxUsableEstimate) {
+            confTarget = maxUsableEstimate;


This seems strange to me. Most of our APIs will already gracefully fail if you pass in too high a conf target, I'd rather we do the same here (same in GUI).

The design of estimateSmartFee is to be used by the wallet to intelligently as possible automatically put a fee on a transaction being sent. In the GUI, the slider clearly indicates what target the fee is being returned for. For the RPC call estimatesmartfee, this is also clearly returned. The difficulty comes with automatic estimation that happens from a sendtoaddress or sendmany RPC call. If in the future we add a way to select a transaction confirm target with the desired RPC call, perhaps it would make sense to fail, but for now it seems to me that its more important for the transaction to go out, even if its paying a fee for a much faster target.

TheBlueMatt · 2017-04-27T23:32:48Z

src/policy/fees.cpp

        fileout << CLIENT_VERSION; // version that wrote the file
        fileout << nBestSeenHeight;
+        if (BlockSpan() > HistoricalBlockSpan()/2) {


This seems like a sharp edge - you can restart a node and end up getting a fee estimate much higher than 30 seconds ago before you shut down. Maybe change the /2 to historicalBest more recent than OLDEST_ESTIMATE_HISTORY*3/4 or so? Would still have the issue but at least its less likely to be hit for those who regularly run a node just during the day or so?

After discussion, I agree that this logic could be more intelligent, but I think that can be left for a later improvement.

instagibbs

utACK 57e71c8 provided conservative is defined to the user

If I find the time I'll do some real-world testing once the buckets fill up.

instagibbs · 2017-05-01T12:54:00Z

src/policy/fees.cpp

@@ -53,13 +55,17 @@ class TxConfirmStats

    double decay;

+    unsigned int scale;


Add description?

instagibbs · 2017-05-01T12:55:07Z

src/policy/fees.cpp

@@ -68,7 +74,8 @@ class TxConfirmStats
     * @param maxConfirms max number of confirms to track
     * @param decay how much to decay the historical moving average per block
     */
-    TxConfirmStats(const std::vector<double>& defaultBuckets, unsigned int maxConfirms, double decay);
+    TxConfirmStats(const std::vector<double>& defaultBuckets, const std::map<double, unsigned int>& defaultBucketMap,


add description above?

instagibbs · 2017-05-01T13:45:34Z

src/policy/fees.h

+    /** Estimate feerate needed to get be included in a block within confTarget
+     *  blocks. If no answer can be given at confTarget, return an estimate at
+     *  the closest target where one can be given.  'conservative' estimates are
+     *  valid over longer time horizons also.


teeny nit: "also" unneeded

instagibbs · 2017-05-01T13:54:04Z

src/policy/fees.cpp

+            confTarget = maxUsableEstimate;
+        }
+
+        assert(confTarget > 0); //estimateCombinedFee and estimateConservativeFee take unsigned ints


I think this is impossible to hit, due to confTarget <= 0 check and maxUsableEstimate <= 1 check. (though this might make sense as belt and suspenders)

instagibbs · 2017-05-01T14:07:12Z

src/rpc/mining.cpp

            "\nEstimates the approximate fee per kilobyte needed for a transaction to begin\n"
            "confirmation within nblocks blocks if possible and return the number of blocks\n"
            "for which the estimate is valid. Uses virtual transaction size as defined\n"
            "in BIP 141 (witness data is discounted).\n"
            "\nArguments:\n"
-            "1. nblocks     (numeric)\n"
+            "1. nblocks       (numeric)\n"
+            "2. conservative  (bool, optional, default=true) Whether to return a more conservative estimate calculated from a longer history\n"


It's not clear to me as the API reader what "conservative" means here, depending on if I'm worrying about over-paying or waiting too long, or being conservative in the amount of data I'm using to estimate.

suggested: "Whether to return a possibly higher feerate estimate calculated from a longer history"

instagibbs · 2017-05-01T14:36:37Z

src/rpc/mining.cpp

+            "  \"decay\" : x.x,          (numeric) exponential decay for historical moving average of confirmation data\n"
+            "  \"scale\" : x,            (numeric) The resolution of confirmation targets at this time horizon\n"
+            "  \"pass.\"                 information about the last range of feerates to succeed in meeting the threshold\n"
+            "  \"fail.\"                 information about the first range of feerates to fail to meet the threshold\n"


nit: s/first/highest/

instagibbs · 2017-05-01T14:36:50Z

src/rpc/mining.cpp

+            "  \"feerate\" : x.x,        (numeric) estimate fee-per-kilobyte (in BTC)\n"
+            "  \"decay\" : x.x,          (numeric) exponential decay for historical moving average of confirmation data\n"
+            "  \"scale\" : x,            (numeric) The resolution of confirmation targets at this time horizon\n"
+            "  \"pass.\"                 information about the last range of feerates to succeed in meeting the threshold\n"


nit: s/last/lowest/

instagibbs · 2017-05-01T14:40:20Z

src/rpc/mining.cpp

+            "  \"startrange\" : x.x,     (numeric) start of feerate range\n"
+            "  \"endrange\" : x.x,       (numeric) end of feerate range\n"
+            "  \"withintarget\" : x.x,   (numeric) number of txs over history horizon in the feerate range that were confirmed within target\n"
+            "  \"totalconfirmed\" : x.x, (numeric) number of txs over history horizon in the feerate range total\n"


s/total/that were confirmed at any point/

instagibbs · 2017-05-01T14:42:42Z

test/functional/smartfees.py

-            # estimatesmartfee should return the same result
-            assert_equal(node.estimatesmartfee(i+1)["feerate"], e)
+            if i >= 13:  # for n>=14 estimatesmartfee(n/2) should be at least as high as estimatefee(n)
+                assert(node.estimatesmartfee((i+1)//2)["feerate"] > float(e) - delta)


could also do a quick check that the transitive property is true for conservative estimates

instagibbs · 2017-05-01T14:53:30Z

src/policy/fees.h

+
+    /** Require an avg of 0.1 tx in the combined feerate bucket per block to have stat significance */
+    static constexpr double SUFFICIENT_FEETXS = 0.1;
+    /** Require an avg of 0.5 tx when using short decay since there are less blocks considered*/


grammar nit: s/less/fewer/

instagibbs · 2017-05-01T15:06:46Z

src/rpc/mining.cpp

+{
+    if (request.fHelp || request.params.size() < 1|| request.params.size() > 3)
+        throw std::runtime_error(
+            "estimaterawfee nblocks\n"


missing threshold

TheBlueMatt · 2017-05-01T15:47:14Z

src/policy/fees.cpp

+            median = doubleEst;
+        }
+
+        if (conservative) {


I believe we may want to add a check that the previous code actually got a result here. And, generally, that all of the four estimate*Fee calls here actually succeed. It seems super strange to accept only one result as fine if the rest fail (eg on startup you may get a response from short horizon on SUFFICIENT_TXS_SHORT, but not on medium/long, also @sdaftuar saw a case this morning where he had an offline node for a while which succeeded in getting a result for conservative but not for non-conservative for a 3 conftarget, I believe because estimateConservativeFee gave a result from the medium/long horizons when the short horizon could not).

After offline discussion, it appears the bigger issue is what to do when the short time horizon estimates have decayed to not being able to give you an answer. It seems like what happens in conservative mode is probably the right thing where you take advantage of the fact that the longer time horizon estimates have not decayed yet. We probably want the ability to utilize this as a fall back in non-conservative mode as well. This can happen either from shutting down a node for a couple of days which means the short time horizon will decay too much, or even just from losing network activity from something like a laptop over the weekend.
As it is, I think its probably still a strict improvement to existing estimates, but could use further refinement.

morcos · 2017-05-03T18:39:51Z

src/policy/fees.cpp

+                if (!conservative) {
+                    // When we aren't using conservative estimates, if a lower confTarget from a more recent horizon returns a lower answer use it.
+                    double medMax = feeStats->EstimateMedianVal(feeStats->GetMaxConfirms(), SUFFICIENT_FEETXS, successThreshold, true, nBestSeenHeight);
+                    if (estimate == -1 ||  medMax < estimate) {


in all 3 of these replacement evaluations (also line 709 and line 722), we probably don't want to do the replacement if the new estimate is -1, should be rare.

morcos · 2017-05-09T17:46:50Z

Addressed the comments in fixup commmits e95eca7 (smarterfee.ver5) and then squashed --> 9a34419 (smarterfee.ver5.squash)

sipa

Code review ACK, with some nits. I haven't reviewed the changes to the estimation algorithm itself.

sipa · 2017-05-09T17:48:14Z

src/policy/fees.cpp

@@ -580,10 +587,15 @@ bool CBlockPolicyEstimator::Write(CAutoFile& fileout) const
 {
    try {
        LOCK(cs_feeEstimator);
-        fileout << 139900; // version required to read: 0.13.99 or later
+        fileout << 149900; // version required to read: 0.14.99 or later


Any reason to keep this number tied to client versions instead of just an independent "fee estimates file version number"?

no reason..., but this also seems simple enough?

sipa · 2017-05-09T17:49:40Z

src/policy/fees.h

@@ -141,19 +207,38 @@ class CBlockPolicyEstimator

    /** Classes to track historical data on transaction confirmations */
    TxConfirmStats* feeStats;
+    TxConfirmStats* shortStats;


Convert these to std::unique_ptrs?

This was the change made in #9942, and I couldn't get unique_ptrs to work while also moving the code to the cpp file. I now also have other ideas how to clean up the wonky design of TxConfirmStats inside CBlockPolicyEstimator, but I don't want to do it now and churn this PR.

sipa · 2017-05-09T18:05:07Z

src/policy/fees.h

@@ -159,6 +171,10 @@ class CBlockPolicyEstimator

 class FeeFilterRounder
 {
+private:
+    static constexpr double MAX_FILTER_FEERATE = 1e7;
+    static constexpr double FEE_FILTER_SPACING = 1.1;


Can you add a comment to explain the relation with the CBlockPolicyEstimator::FEE_SPACE constant?

yeah, there is no relation.
when I made the filter spacing, I just copied the other one arbitrarily, but then I decided if I was changing the estimate spacing, there was no reason to also make a change to the filter spacing. I can add a comment.

sipa · 2017-05-09T18:14:03Z

src/rpc/mining.cpp

+            "\nWARNING: This interface is unstable and may disappear or change!\n"
+            "\nWARNING: This is an advanced API call that is tightly coupled to the specific\n"
+            "         implementation of fee estimation. The parameters it can be called with\n"
+            "         and the results it returns will change if the internal implementaion changes.\n"


@jlopp's comment hasn't been addressed yet: "nit: Implementation".

sipa · 2017-05-09T18:16:47Z

src/rpc/mining.cpp

+        if (horizonInt < 0 || horizonInt > 2) {
+            throw JSONRPCError(RPC_TYPE_ERROR, "Invalid horizon for fee estimates");
+        }
+        else {


Style nit: else on the same line as }

sipa · 2017-05-09T18:18:16Z

src/rpc/mining.cpp

+            "2. threshold   (numeric, optional) The proportion of transactions in a given feerate range that must have been\n"
+            "               confirmed within nblocks in order to consider those feerates as high enough and proceed to check\n"
+            "               lower buckets.  Default: 0.95\n"
+            "3. horizon     (numeric, optional) How long a history of estimates to consider. 0=short, 1=medium, 2=long.\n"


To make the RPC interface less dependent on the specific implementation, perhaps don't make horizon a parameter, and just return information for all applicable horizons for the given value of nblocks?

slightly prefer existing, but I could change it if prevailing opinion is otherwise..
estimaterawfee is meant to be implementation dependent

The reason for asking is that while the output of this RPC may be implementation dependent, it would be nice if (to the extent possible), its arguments aren't. The specific buckets used seem much more likely to change in other revisions.

Just a nit.

sipa · 2017-05-09T18:25:15Z

src/policy/fees.cpp

+
+double CBlockPolicyEstimator::estimateCombinedFee(unsigned int confTarget, double successThreshold, bool conservative) const
+{
+    double estimate = -1;


Can you write this function as a loop over the 3 horizons, instead of a cascaded if/then/else?

I can take a pass at rewriting the logic to be a bit clearer, but I'm not sure a loop is the right thing. There are basically 6 different combinations of estimates you might want to check depending on what range your target is in and whether or not you want to check shorter horizons. But the logic might be cleaner if those are just spelled out.

sipa · 2017-05-09T18:27:34Z

src/policy/fees.h

@@ -173,6 +173,10 @@ class CBlockPolicyEstimator
 private:
    CFeeRate minTrackedFee;    //!< Passed to constructor to avoid dependency on main
    unsigned int nBestSeenHeight;
+    unsigned int firstRecordedHeight;
+    unsigned int historicalFirst;


Nit: unused until 2 commits later.

I wanted to get the file format change over and done with...

sipa · 2017-05-09T18:38:01Z

src/policy/fees.cpp

+unsigned int CBlockPolicyEstimator::MaxUsableEstimate() const
+{
+    // Block spans are divided by 2 to make sure there are enough potential failing data points for the estimate
+    return std::min(longStats->GetMaxConfirms(), std::max(BlockSpan(), HistoricalBlockSpan()) / 2);


Nit: If you'd return which of the two BlockSpans was chosen in a pair or a bool, you wouldn't need to repeat the decision logic in the debug print statement of processBlock.

after actually doing this, it seemed uglier, so leaving it as is

…zons. Make feerate buckets smaller (5% instead of 10%) and make the 3 different horizons have half lifes of 3 hours, 1 day and 1 week respectively.

Instead of stopping if it encounters a "sufficient" number of transactions which don't meet the threshold for being confirmed within the target, it keeps looking to add more transactions to see if there is a temporary blip in the data. This allows a smaller number of required data points.

…update curNearBucket on final loop.

morcos · 2017-05-17T19:48:22Z

Added an extra commit to improve JSON output as per @sipa. Was kind of annoying to squash, so will leave it if that's ok?

@laanwj I think this is reviewed enough that we'd be better off merging and getting some more experimentation with the estimates.. It takes 2 full weeks for the longest estimates to even be exposed.

The biggest open questions are whether I should add an ability to use old estimates during transition and whether we are going to build in GUI support.

The tests could also use some improvement but I don't think they're notably worse than they were.

sipa · 2017-05-17T19:58:56Z

utACK 38bc1ec

instagibbs · 2017-05-17T20:01:16Z

utACK 38bc1ec

38bc1ec Make more json-like output from estimaterawfee (Alex Morcos) 2d2e170 Comments and improved documentation (Alex Morcos) ef589f8 minor cleanup: remove unnecessary variable (Alex Morcos) 3ee76d6 Introduce a scale factor (Alex Morcos) 5f1f0c6 Historical block span (Alex Morcos) aa19b8e Clean up fee estimate debug printing (Alex Morcos) 10f7cbd Track first recorded height (Alex Morcos) 3810e97 Rewrite estimateSmartFee (Alex Morcos) c7447ec Track failures in fee estimation. (Alex Morcos) 4186d3f Expose estimaterawfee (Alex Morcos) 2681153 minor refactor: explicitly track start of new bucket range and don't update curNearBucket on final loop. (Alex Morcos) 1ba43cc Make EstimateMedianVal smarter about small failures. (Alex Morcos) d3e30bc Refactor to update moving average on fly (Alex Morcos) e5007ba Change parameters for fee estimation and estimates on all 3 time horizons. (Alex Morcos) c0a273f Change file format for fee estimates. (Alex Morcos) Tree-SHA512: 186e7508d86a1f351bb656edcd84ee9091f5f2706331eda9ee29da9c8eb5bf67b8c1f2abf6662835560e7f613b1377099054f20767f41ddcdbc89c4f9e78946d

RHavar · 2017-05-18T00:18:43Z

I tried reading the description from help, and I don't really understand this:

[the rpc command also returns] the number of blocks for which the estimate is valid.

but from the actual return:

"blocks" : n (numeric) block number where estimate was found

Without reading the code, it's not clear what blocks means. Is it the block height when you called estimatestartfee or is it how long it's valid for?

instagibbs · 2017-05-18T02:14:13Z

@RHavar without actually looking at the code, I believe this refers to the "cheapest bucket" block which that feerate was found.

RHavar · 2017-05-18T02:39:39Z

@instagibbs Yeah, you seem right. But I don't think that was inferable from the help information :P

What's the motivation for estimatefee returning a different value than estimatesmartfee ?

$ bitcoin-cli estimatefee 6
0.00254275

$ bitcoin-cli estimatesmartfee 6
{
  "feerate": 0.00208824,
  "blocks": 6
}

?

Anyway, fantastic work. This is a huge improvement, something that has been long needed. Well done @morcos

RHavar · 2017-05-18T02:55:06Z

I've updated https://estimatefee.com/ to now use estimatesmartfee $n so if you're interested you can take a look there.

The bitcoin node has only been up a few hours, so it'll probably take a while to collect enough information.

P.S. you kinda broke my domain name lol

morcos · 2017-05-18T17:08:50Z

@RHavar you might be right that the description is not the clearest.
blocks refers to the target for which you are getting an estimate, which is an expected number of blocks within which you will be confirmed. Occasionally you ask for an estimate at one target but it returns you an estimate for a different target (b/c it is unable to give you an estimate for the target you asked for)

RHavar · 2017-05-18T18:42:40Z

@morcos In that case I'd suggest renaming it from "blocks" to "target" and changing the description to what you wrote above :P

karelbilek · 2017-09-02T20:29:58Z

Not sure where to put it, but just FYI

I made a simple tool that looks up smart fees once per 10 minutes and puts it here

https://estimatesmartfee.com/html.html

or

https://estimatesmartfee.com/json.json

If you want to use it for testing or something

jb55 · 2017-10-26T06:37:52Z

also FYIs, I put together a small bash script that uses this for quickly grabbing fees from the command line: https://bitcoin.stackexchange.com/a/61348/1235

thanks!

38bc1ec Make more json-like output from estimaterawfee (Alex Morcos) 2d2e170 Comments and improved documentation (Alex Morcos) ef589f8 minor cleanup: remove unnecessary variable (Alex Morcos) 3ee76d6 Introduce a scale factor (Alex Morcos) 5f1f0c6 Historical block span (Alex Morcos) aa19b8e Clean up fee estimate debug printing (Alex Morcos) 10f7cbd Track first recorded height (Alex Morcos) 3810e97 Rewrite estimateSmartFee (Alex Morcos) c7447ec Track failures in fee estimation. (Alex Morcos) 4186d3f Expose estimaterawfee (Alex Morcos) 2681153 minor refactor: explicitly track start of new bucket range and don't update curNearBucket on final loop. (Alex Morcos) 1ba43cc Make EstimateMedianVal smarter about small failures. (Alex Morcos) d3e30bc Refactor to update moving average on fly (Alex Morcos) e5007ba Change parameters for fee estimation and estimates on all 3 time horizons. (Alex Morcos) c0a273f Change file format for fee estimates. (Alex Morcos) Tree-SHA512: 186e7508d86a1f351bb656edcd84ee9091f5f2706331eda9ee29da9c8eb5bf67b8c1f2abf6662835560e7f613b1377099054f20767f41ddcdbc89c4f9e78946d

fanquake added the TX fees and policy label Apr 12, 2017

jlopp reviewed Apr 19, 2017

View reviewed changes

Change file format for fee estimates.

c0a273f

Move buckets and bucketMap to be stored as part of overall serialization of estimator. Add some placeholder data so file format is only changed once. Maintain 3 different TxConfirmStats with potential for different decays and scales.

morcos force-pushed the smarterfee branch from 7306590 to 57e71c8 Compare April 20, 2017 20:09

laanwj added this to the 0.15.0 milestone Apr 20, 2017

morcos mentioned this pull request Apr 26, 2017

Always log debug information for fee calculation in CreateTransaction #10284

Merged

TheBlueMatt reviewed Apr 28, 2017

View reviewed changes

instagibbs approved these changes May 1, 2017

View reviewed changes

instagibbs reviewed May 1, 2017

View reviewed changes

TheBlueMatt reviewed May 1, 2017

View reviewed changes

morcos commented May 3, 2017

View reviewed changes

morcos force-pushed the smarterfee branch from e95eca7 to 9a34419 Compare May 9, 2017 17:41

sipa reviewed May 9, 2017

View reviewed changes

sdaftuar mentioned this pull request May 10, 2017

Rename TxConfirmStats to CTxConfirmStats to achieve class naming consistency #10378

Closed

morcos added 4 commits May 10, 2017 11:45

Change parameters for fee estimation and estimates on all 3 time hori…

e5007ba

…zons. Make feerate buckets smaller (5% instead of 10%) and make the 3 different horizons have half lifes of 3 hours, 1 day and 1 week respectively.

Refactor to update moving average on fly

d3e30bc

minor refactor: explicitly track start of new bucket range and don't …

2681153

…update curNearBucket on final loop.

sipa merged commit 38bc1ec into bitcoin:master May 17, 2017

TheBlueMatt mentioned this pull request May 17, 2017

Need fee estimation migration for 0.15 #10418

Closed

practicalswift mentioned this pull request May 18, 2017

[trivial] Fix three recently introduced typos #10419

Merged

EagleTM mentioned this pull request May 18, 2017

Prepare to use new estimatesmartfee in client spesmilo/electrum#2420

Closed

ryanofsky mentioned this pull request May 22, 2017

Add fee_est tool for debugging fee estimation code #10443

Closed

jnewbery mentioned this pull request Aug 1, 2017

TODO for release notes 0.15.0 #9889

Closed

12 tasks

This was referenced Oct 5, 2017

Backport smart fees from bitcoin core Bitcoin-ABC/bitcoin-abc#94

Closed

[feature request] Backport smart fees from bitcoin core litecoin-project/litecoin#385

Closed

[feature request] Backport smart fees from bitcoin core dashpay/dash#1669

Closed

jonasschnelli mentioned this pull request Oct 15, 2017

estimatesmartfee always returns "Insufficient data or no feerate found" #11500

Closed

brunocalmels mentioned this pull request Aug 1, 2021

Fees estimate enhancements bcoin-org/bcoin#1022

Open

bitcoin locked as resolved and limited conversation to collaborators Dec 16, 2021

		@@ -53,13 +55,17 @@ class TxConfirmStats

		double decay;

		unsigned int scale;

Better fee estimates #10199

Better fee estimates #10199

Uh oh!

Conversation

morcos commented Apr 12, 2017 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

jonasschnelli commented Apr 13, 2017

Uh oh!

jonasschnelli commented Apr 19, 2017

Uh oh!

Choose a reason for hiding this comment

Uh oh!

jlopp commented Apr 19, 2017

Uh oh!

sipa commented Apr 19, 2017 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

morcos commented Apr 20, 2017

Uh oh!

crwatkins commented Apr 24, 2017

Uh oh!

jlopp commented Apr 25, 2017 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

morcos commented Apr 25, 2017

Uh oh!

crwatkins commented Apr 25, 2017

Uh oh!

Leviathn commented Apr 25, 2017

Uh oh!

TheBlueMatt left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

instagibbs left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

morcos commented Apr 12, 2017 •

edited

Loading

sipa commented Apr 19, 2017 •

edited

Loading

jlopp commented Apr 25, 2017 •

edited

Loading