Research · Network Economics
Bitcoin Metcalfe Value Monitor
Price against a network valuation using a Metcalfe-style specification: value scales with the square of network size, adjusted for coin supply. Network size is proxied by a running total of daily active addresses, in which an address is counted again on every day it is used: a measure of accumulated activity, not of how many people use bitcoin. Fitted live in your browser from public on-chain data. A descriptive lens, not a fair-value engine: since 2017 calendar time explains price about as well.
Current readings
Computed from the daily public snapshot for 14 October 2026, published 2026-10-15 11:22 UTC. The charts and tables below are drawn in the browser from the same file; these figures are the ones they show.
| Metcalfe value, validated fit from 2017 | $78,094 | cumulative address-activity proxy, n², no lost-coin adjustment, fit 2017 onward |
|---|---|---|
| Price against it | +6.4% | 6.2 log points; 56th percentile |
| Out-of-sample error | 50.1 log points | one year ahead, annual refits |
| Fit from 2011, for comparison | $138,792 | matches published figures; out-of-sample error 148 log points, fails validation |
| Network-size proxy | 2,403,654,330 | running total of daily address activity |
Where price sits relative to the network-value estimate
The gap between price and the model, placed on the fit window's own history. Regime cuts are the 10th, 30th, 70th and 90th percentiles of that history, which stay meaningful when bubbles skew it. The percentile matters more than whether price is above or below the line.
Price and Metcalfe value
Price against the model
Model fit
- Proxy, nCumulative address activity
- ModelMetcalfe, exponent fixed at 2
- Fit window2017-01-01 to 2026-10-14 (3,574 days)
- R² of log price, fit window0.841
- R² of log price, full history0.776
- Residual skewness, fit window0.19 (near symmetric)
- Residual excess kurtosis-0.71
- Network exponent2, fixed by the model
- Lost coinsnone
- Effective supply today20.10M BTC of 20.10M mined
- Scale constant k2.72e-7
- Out-of-sample error, annual refits50 pts RMSE, bias +12 pts, 2479 predicted days
- Today's value across those refits$65,218 to $78,173
- Residual σ47.9 log points
- Furthest above value+321% on Dec 17, 2017
- Furthest below value-64% on Mar 13, 2020
How far from the model
- Days above Metcalfe value49% of history
- Median gap-1.2%
- 10th to 90th percentile-46% to +82%
- 30th to 70th percentile-28% to +33%
- Current state, days269
What followed, by starting gap
| Price against the model | 90d | 180d | 365d | Days in bucket |
|---|---|---|---|---|
| more than 50% below | +62% | +165% | +161% | 288 |
| 25% to 50% below | +16% | +46% | +135% | 899 |
| up to 25% below | +16% | +12% | +82% | 631 |
| up to 25% above | +2% | +26% | +63% | 500 |
| 25% to 50% above | -12% | -13% | -16% | 411 |
| 50% to 100% above | -2% | -15% | -35% | 596 |
| more than double | -32% | -26% | -64% | 249 |
Descriptive, overlapping windows, in-sample. Not a forecast. With a fit window that starts before 2014, the rows are dominated by the 2011 to 2013 bubbles and the medians run to four figures.
Out-of-sample validation
| Model | Fit window | Error, RMSE | Bias | Today's value across refits | Refits |
|---|---|---|---|---|---|
| n², exponent 2 | Full history | 161 | -135 | $158,089 to $1,472,761 | 13 |
| n², exponent 2 | Since 2011 | 148 | -120 | $143,123 to $1,181,617 | 13 |
| n², exponent 2 | Since 2012 | 114 | -82 | $109,906 to $541,918 | 12 |
| n², exponent 2 | Since 2014 | 53 | -12 | $80,419 to $106,331 | 10 |
| n², exponent 2 | Since 2017 | 50 | +12 | $65,218 to $78,173 | 7 |
| Fitted exponent | Full history | 90 | +45 | $10,439 to $44,873 | 13 |
| Fitted exponent | Since 2011 | 92 | +50 | $8,194 to $45,640 | 13 |
| Fitted exponent | Since 2012 | 97 | +19 | $10,711 to $541,096 | 12 |
| Fitted exponent | Since 2014 | 99 | +70 | $454.66 to $63,732 | 10 |
| Fitted exponent | Since 2017 | 58 | +5 | $49,604 to $129,706 | 7 |
| Activity adjusted | Full history | 93 | +48 | $5,591 to $38,440 | 13 |
| Activity adjusted | Since 2011 | 94 | +49 | $3,756 to $37,340 | 13 |
| Activity adjusted | Since 2012 | 96 | +14 | $10,272 to $526,629 | 12 |
| Activity adjusted | Since 2014 | 98 | +52 | $161.24 to $38,850 | 10 |
| Activity adjusted | Since 2017 | 59 | +11 | $33,767 to $82,595 | 7 |
Lowest error: n², exponent 2, since 2017 (50 pts). Current proxy: Cumulative address activity.
Is it the network, or just time?
The address total behind this page is a running sum of a daily count that has been roughly flat since 2017, so it rises almost exactly like time: since 2017 its logarithm correlates 0.994 with the logarithm of bitcoin’s age. A model of price on time alone, with no network data, does about as well one year ahead: 60 log points of error against 58 for the network model. The network model was better in 5 of 7 refit years, by 2.7 points pooled, and adds 0.016 to R² beyond time. The fit on this page is real; it is not evidence that network size, rather than the passage of time, explains price.
| Price explained by | In-sample R² | Out-of-sample error |
|---|---|---|
| Time alone (days since 3 January 2009) | 0.832 | 60 log points |
| Network alone (n² / supply) | 0.847 | 58 log points |
| Both together | 0.848 | 77 log points |
The years scored here start in 2020, a little later than the validation table above, which is why the network model's error differs slightly from it. The joint model predicts worst because the two inputs move almost together and the fit cannot tell them apart. Descriptive, not causal: address activity responds to price as well as driving it.
External references
| Estimate | Value | As of | Price versus it | Versus this page |
|---|---|---|---|---|
| This page, current specification | $78,095 | 2026-10-14 | +6% | |
| This page, fit from 2011 (fails validation) | $138,792 | 2026-10-14 | -40% | +78% |
| Cane Island, Metcalfe value (MET) | $138,557 | 2026-09-01 | -40% | +77% |
| Cane Island, Macro model value (MAC) | $109,000 | 2026-08-31 | -24% | +40% |
Cane Island's methodology is proprietary. Fitting this page from 2011, with no lost-coin adjustment, reproduces their MET within a few percent, which is why that row is shown. The same fit fails this page's own out-of-sample test, overvaluing the following year by about three times on average, so it is not the default. The validated 2017 fit gives a value a little over half as large; their MAC applies a macro overlay that moves in the same direction.
Specification comparison
| Proxy, n | Model | Value today | Price against it | R² window | Exponent | Out-of-sample error | Status |
|---|---|---|---|---|---|---|---|
| Cumulative address activity | n², exponent 2 | $78,095 | +6% | 0.841 | 2.00 | 50 pts, refits $65,218 to $78,173 | validated |
| Cumulative address activity | Fitted exponent (diagnostic) | $87,886 | -5% | 0.847 | 2.18 | 58 pts, refits $49,604 to $129,706 | weaker |
| Cumulative address activity | Activity adjusted (diagnostic) | $79,394 | +5% | 0.849 | 2.27 | 59 pts, refits $33,767 to $82,595 | weaker |
| Addresses with balance | n², exponent 2 | $50,592 | +64% | 0.766 | 2.00 | 69 pts, refits $30,424 to $48,868 | weaker |
| Addresses with balance | Fitted exponent (diagnostic) | $84,465 | -2% | 0.887 | 3.17 | 56 pts, refits $71,382 to $154,628 | weaker |
| Addresses with balance | Activity adjusted (diagnostic) | $73,010 | +14% | 0.891 | 3.35 | 57 pts, refits $41,823 to $81,560 | weaker |
| Addresses, 8y half-life | n², exponent 2 | $54,114 | +54% | 0.793 | 2.00 | 61 pts, refits $37,332 to $52,622 | weaker |
| Addresses, 8y half-life | Fitted exponent (diagnostic) | $73,833 | +13% | 0.845 | 2.66 | 58 pts, refits $38,930 to $91,009 | weaker |
| Addresses, 8y half-life | Activity adjusted (diagnostic) | $72,130 | +15% | 0.845 | 2.68 | 60 pts, refits $26,599 to $71,390 | weaker |
| Gompertz trend | n², exponent 2 | $56,483 | +47% | 0.793 | 2.00 | 59 pts, refits $40,436 to $55,087 | weaker |
| Gompertz trend | Fitted exponent (diagnostic) | $71,068 | +17% | 0.822 | 2.47 | 61 pts, refits $28,401 to $70,114 | weaker |
| Gompertz trend | Activity adjusted (diagnostic) | $74,087 | +12% | 0.822 | 2.43 | 64 pts, refits $18,218 to $73,331 | weaker |
| Unspent outputs | n², exponent 2 | $68,304 | +22% | 0.729 | 2.00 | 69 pts, refits $42,363 to $67,948 | weaker |
| Unspent outputs | Fitted exponent (diagnostic) | $81,452 | +2% | 0.742 | 2.30 | 100 pts, refits $80,189 to $3,633,768 | fails |
| Unspent outputs | Activity adjusted (diagnostic) | $47,802 | +74% | 0.788 | 3.07 | 91 pts, refits $42,060 to $2,628,512 | fails |
Fitted-exponent and activity-adjusted rows are diagnostics: they show what exponent the data prefers and whether transaction activity adds information. They are not offered as valuations. Status compares each row's out-of-sample error with the best row.
Data audit
| Series | First | Last | Days | Min | Max | Latest |
|---|---|---|---|---|---|---|
| Price, USD | 2010-08-18 | 2026-10-14 | 5,902 | $0.0600 | $124,777 | $83,094 |
| Daily unique addresses | 2010-08-18 | 2026-10-14 | 5,902 | 290 | 1,072,862 | 480,301 |
| Cum. address activity | 2010-08-18 | 2026-10-14 | 5,902 | 106,402 | 2.40B | 2.40B |
| Circulating supply, BTC | 2010-08-18 | 2026-10-14 | 5,902 | 3,752,500 | 20,098,369 | 20,098,369 |
| Transactions per day | 2010-08-18 | 2026-10-14 | 5,902 | 260 | 927,010 | 803,842 |
| Addresses with a balance | 2010-08-18 | 2026-10-14 | 5,902 | 54,907 | 57.2M | 57.2M |
| Unspent outputs | 2010-08-18 | 2026-10-14 | 5,902 | 58,069 | 189.3M | 166.1M |
- Daily calendar continuousyes
- Supply never decreasesyes
- All prices positiveyes
- Address feed gaps filled by interpolation31 days between 2009-01-03 and 2026-08-24, linear between observed points
- Value recomputed from k, n, effective S at last day$78,095 matches chart
- Mean log residual in fit window-1.6e-13 pts (zero, as required by least squares)
- Premium at last close recomputed6.20 vs series 6.20 log points
- Data sourcedaily snapshot
- Live price vs last daily closeno live price
- Percentile is of fit-window premiums3,574 days from 2017-01-01
Network-size proxy
Method
Following Peterson (2018), Bitcoin is treated as a token currency whose network value is proportional to the square of the number of users, while supply dilution is handled by dividing by coins in circulation. The per-coin Metcalfe value is MET = k × n² / S, where S is total bitcoin in circulation and k is fitted by least squares on log price over the chosen fit window.
The choice of n matters more than anything else on this page. Peterson's demand variable was the number of wallets, a stock that only grows. A running total of Blockchain.com's daily unique addresses since 2009 behaves like that stock and yields a network exponent close to two. It is not a count of addresses ever used: an address active on a hundred days is counted a hundred times, which is why the total passes two billion while about 57 million addresses hold a balance. Daily unique addresses are a flow, not a count of unique people or users: they run at roughly half a million a day, about where they were in 2017 to 2019, so a model built on them returns that era's price level (roughly $6,000 to $23,000 depending on window) with an R² near zero since 2017 and out-of-sample error above 140 log points. That definition was tested and rejected, and it is deliberately not offered on this page. Users also leave, so a decay option is provided: each day's unique addresses join the user stock and then decay exponentially with a chosen half-life. A half-life of zero collapses to the daily flow and an infinite half-life to the raw cumulative count. On this data the fit improves monotonically as the half-life lengthens, and the undecayed total fits the 2014 onward era best, which is why it is the default proxy; "Best fit" runs that comparison and reports the winner. Every year of half-life removed lowers the Metcalfe value, so decay is a conservative adjustment, not a neutral one. When the daily snapshot is available, the count of addresses holding a non-zero balance (Coin Metrics) is offered as well; it is the adoption proxy used in the published power-law decomposition and the closest public analogue of a wallet count. Unspent outputs are a further proxy for holders. The Gompertz option follows Peterson (2019): the user series is replaced by a fitted sigmoid growth curve, so the Metcalfe value becomes a smooth curve and the premium series isolates what the paper calls non-economic deviations, the 2011, 2013 and 2017 episodes among them.
Lost coins enter through supply: S in the formula is the surviving supply, not the mined total. The reader chooses the total assumed lost today, from the published range (Chainalysis 2.3M to 3.7M, later estimates near 3.8M, a 5.6M ceiling for coins dormant ten years or more, most analysts centring on 3 to 4 million), and a loss profile. The page then calibrates a per-year loss hazard on each vintage so that the profile reaches that total, and rebuilds the surviving supply day by day. Two facts govern how much this matters. A share of supply lost uniformly throughout the fit window changes nothing, because the constant k absorbs it; only the change in the lost share across the window moves the value. And the profile is what sets that change: coins were overwhelmingly lost in the early years, when they were mined for cents and discarded, and rarely today, when supply sits with custodians. The profile offered by default is front-loaded (a hazard that halves every three years above a small floor); it lifts the value about 2% at the central estimate on the 2017 window and about 4% on the 2011 window. The adjustment is off by default because published figures do not state one. A constant-rate profile is offered for comparison and lifts it about 7%, because it assumes losses continue at the early rate. The better input would be measured dormancy from the chain (supply last active ten or more years ago), which is not in the public feed used here. Two diagnostic forms appear in the comparison tables but are not offered as valuations. The fitted-exponent form lets the data choose the network exponent instead of fixing it at two; on the default window it returns an exponent close to two, which supports the law, but as a valuation it is unstable across annual refits. The activity-adjusted form adds the 30-day average transaction count as a second regressor; it fits marginally better in sample and predicts worse out of sample. A macro overlay in the spirit of Cane Island's Macro Model Value, adding the broad trade-weighted dollar index and the effective federal funds rate as regressors, was tested the same way on monthly data: it raised in-sample R² from 0.85 to 0.93 on the 2014 window and raised out-of-sample error from 55 to 69 points, with the value for today swinging across a six-fold range between annual refits. It is not offered here. Nor are price-only constructions such as power-law floors in calendar time or seasonal-year profiles: they contain no user count, and this page is a network valuation. The Gompertz option is the counterpart of adoption-curve estimates, and it is kept as a diagnostic because it predicts less well than the observed user stock.
The premium series is 100 × ln(price / MET), in log points: 50 log points above is price 65% above the model, 50 below is 39% below, and +622 is about 500 times. The page labels every gap as the real percentage and keeps log points for the statistics. Standard units divide the log-point series by the standard deviation of residuals in the fit window. Regimes are cut at the 10th, 30th, 70th and 90th percentiles of the fit-window premium rather than at multiples of σ, because the residuals are right-skewed: bubbles push price far above value for months, while discounts are shallower and longer. All distribution statistics (percentile, extremes, regime shares, forward returns) are computed over the fit window; earlier years are drawn on the charts, shaded as out of sample, but not fitted.
Validation. Every model and window is also scored out of sample: fit on data through each 31 December from the window start, predict the following year, and repeat. On this data the fixed-exponent model on the running address total, fitted from 2014 or 2017, predicts with about 50 log points of error and small bias; fitted from 2017, its value for today moves by about ±10% across annual refits, and from 2014 by about ±15%. The fitted-exponent and activity-adjusted forms fit better in sample but predict worse, and their value for today swings by a factor of ten across refits, because the extra parameter is not identified by a series that grows six orders of magnitude. The full-history and 2011 fits overvalue systematically, by about three to four times on average in the year after each refit. These results, not in-sample R², decide the defaults, and the dashed line on the main chart shows what the chosen method would have said on each date. A separate test, shown above, compares the model with calendar time alone, and finds them about equal since 2017.
The full-history calibration is a steep function of its start date: the first months of exchange trading in 2010 fit the law worst by a factor of fifty and pull the constant up. Starting the fit on 1 January 2011 removes that distortion and reproduces the published Cane Island Metcalfe value within a few percent, so it is kept one click away and in the references table. It is not the default: it fails the out-of-sample test. The 2017 window, covering three halving cycles, has the lowest out-of-sample error and is the default. Before 2014 the address stock is still forming, so n² understates the young network's value by orders of magnitude; fitting through those years inflates the residual σ and, because the early residuals are positive, pulls the constant k upward. A 2014 start keeps the 2017 bubble at the front of the window and leaves a strongly right-skewed residual. From 2017 onward the specification comparison converges: all four model forms on the cumulative user series land within about ±10% of each other, the fitted exponent is close to two, and residuals are near symmetric. Earlier windows are left available for comparison.
What this page tells you, and what it does not
Descriptive: the relationship between the network-size proxy and price has held on the mature network, and its fit statistics are printed. Forecasting over a defined horizon: the page tests one-year-ahead prediction with annual refits and prints the error and bias; that is the strongest predictive claim it makes. Not claimed: a price target, or causation. Address counts respond to price as well as drive it, since bull markets bring users on chain and bear markets send them away; a cumulative stock damps that feedback but does not remove it. The model contains no explicit calendar term, but its network input rises almost exactly like time, and since 2017 a model of price on time alone predicts about as well (see "Is it the network, or just time?"). Address counts are a weak proxy for economic network size: one person holds many addresses, while exchanges, custodians and ETFs hold the coins of millions of people in a few, so the proxy undercounts recent adoption. Metcalfe's law is a heuristic, not an established law for bitcoin, whose price also reflects scarcity, liquidity, institutional demand and speculation. Treat a large gap as something to investigate beside capital flows, holder behaviour and liquidity, not as a buy or sell signal, and size positions as if the stated error is real.
Sources
Unique addresses, transactions, circulating supply and daily price from Blockchain.com, read from the site's daily public snapshot (btc-data), with the Blockchain.com Charts API as a fallback. Addresses with a balance from Coin Metrics, through the same snapshot. Spot price from Coinbase, refreshed every minute while the page is open. Data is cached in this browser for six hours.
References: T. F. Peterson, "Metcalfe's Law as a Model for Bitcoin's Value," Alternative Investment Analyst Review, 2018, SSRN 3078248; T. F. Peterson, "Bitcoin Spreads Like a Virus," 2019, SSRN 3356098. Peterson writes value as pairwise connections, n(n-1)/2; that is indistinguishable from n² here, the factor of one half being absorbed into k.
Companion page: the Bitcoin Power Law Monitor, where the Metcalfe exponent is one half of the identity β = βA × βM that decomposes the price power law in time.
Metcalfe value measures demand-driven network effects against a fixed-supply asset. It is not a price forecast and nothing here is investment advice. Address counts are an imperfect proxy for users and are affected by exchange batching, ordinals, and privacy tooling.