Methodology

This page describes how the published factors are built. Why we build them this way is explained on The Standard.

Data

We use two survivorship-bias-free sources: CoinGecko (canonical) and CoinMarketCap. Both are retrieved with the crypto2 R package (CRAN), which keeps delisted and inactive coins. We store daily price, market capitalisation and volume of every coin, and daily open/high/low/close where available.

One coin, one id. Every coin gets our own identifier (uid, e.g. OCP004106 for Bitcoin). It links the CoinMarketCap id and the CoinGecko id. Each candidate match is proposed by evidence (contract address, website/X/GitHub links, name and symbol, slug, DefiLlama) and is only accepted if the two providers’ prices agree. Identifiers are never reused. The crosswalk is published under CC BY 4.0 (see Data).

Dates. All data use end-of-day UTC labels: the price for date dd is the close at 24:00 UTC on dd. CoinGecko labels its daily data at the start of the day, and so did CoinMarketCap before 1 February 2018; we shift those labels by one day. CoinMarketCap’s archive has no separate value for 31 January 2018 because of that switch. There, and only there, one daily return spans two days.

Cleaning

Every series goes through the same baseline cleaning, following Fieberg et al. (2024):

  1. keep coin-days with positive price, market cap and volume;
  2. drop coin-days with a market cap above Bitcoin’s on the same day (Fieberg et al., 2025);
  3. reversal filter (Ince & Porter, 2006): if a daily return is at least 300% and the two-day return is below 50%, both daily returns are set to missing;
  4. daily returns above 1,000% are set to missing;
  5. a weekly return requires a complete run of daily returns in that week.

Excluded from the start: stablecoins, derived tokens (wrapped, bridged, staked and receipt tokens such as WBTC, stETH or Compound/Venus tokens) and tokenised traditional assets (stocks, ETFs, gold, Treasury bills). Derived tokens and tokenised assets are claims on other assets, so their returns are not crypto-market returns.

The standard factors

Let ℰt\mathcal{E}_t be the eligible coins in week tt: coins with a weekly return and a previous-week average market cap mi,t−1≥$1m_{i,t-1} \geq \$1M. Weights are wi,t=mi,t−1/∑jmj,t−1w_{i,t} = m_{i,t-1} / \sum_{j} m_{j,t-1} within each portfolio.

Weeks. ISO weeks: Monday 00:00 to Sunday 24:00 UTC, so a weekly return runs from Sunday close to Sunday close. week_start is the Monday and week_end the Sunday.

Breakpoints. Tercile breakpoints are computed only from coins with mi,t−1≥$100m_{i,t-1} \geq \$100M. All eligible coins are then assigned to the three portfolios, in analogy to NYSE breakpoints in equity markets (Fama & French, 1993). A leg needs at least five coins; otherwise the factor is missing that week.

CMKT, market. The value-weighted return of all eligible coins in excess of the risk-free rate: CMKTt=∑i∈ℰtwi,tri,t−RFt.\text{CMKT}_t = \sum_{i \in \mathcal{E}_t} w_{i,t}\, r_{i,t} - RF_t . RFtRF_t is the 3-month US T-bill rate (FRED series DTB3). Crypto trades on 365 days a year, so the rate is compounded over every calendar day of the week.

CSIZE, size. Small minus big: the value-weighted return of the bottom tercile of mi,t−1m_{i,t-1} minus that of the top tercile.

CMOM, momentum. Winners minus losers: terciles of the return over the previous two weeks, r2,0=Pi,t−1Pi,t−3−1,r_{2,0} = \frac{P_{i,t-1}}{P_{i,t-3}} - 1 , where Pi,tP_{i,t} is the last close of week tt. The coin must be present in both weeks. Each portfolio is value-weighted.

Choice Standard
Source CoinGecko (canonical); CoinMarketCap alongside
Weeks ISO weeks, close Sunday 24:00 UTC
Eligibility previous-week average cap ≥ $1M; no age or price screen
Exclusions stablecoins, derived tokens, tokenised traditional assets
Weights value (previous-week average cap)
Sorts terciles; breakpoints from coins ≥ $100M; ≥ 5 coins per leg
Momentum signal r2,0r_{2,0}, the return of the previous two weeks
Winsorisation none
Risk-free rate 3-month T-bill (DTB3), 365-day accrual
Implementation lag none (a 1-day-lag version is published)

Published variants

  • 1-day lag (_lag1): each weekly window starts and ends one day later than the window used for sorting (Monday to Monday close). This leaves a day to trade on the signal.
  • All-coin breakpoints (_all): identical, but breakpoints are computed from all eligible coins. We publish it for comparison only; see The Standard for why it is not a fair benchmark.
  • Other calendars: the standard on 7-day weeks closing on each weekday, and on the 52-week calendar of Liu et al. (2022) (January 1–7 = week 1; week 52 absorbs the last 8–9 days of the year).

Liu, Tsyvinski & Wu (2022): exact construction

As a replication anchor we also publish the factors exactly as defined by Liu et al. (2022), on our cleaned data:

  • coins with a previous end-of-week market cap of at least $1M, value-weighted, stablecoins included, no winsorisation;
  • CSMB (our CSIZE column): the bottom 30% minus the top 30% by market cap;
  • CMOM: an independent double sort on size (30/40/30, neutral size dropped) and three-week momentum r3,0=Pi,t−1/Pi,t−4−1r_{3,0} = P_{i,t-1}/P_{i,t-4} - 1 (30/40/30): CMOMt=12(RSW,t+RBW,t)−12(RSL,t+RBL,t).\text{CMOM}_t = \tfrac12\,(R_{SW,t} + R_{BW,t}) - \tfrac12\,(R_{SL,t} + R_{BL,t}) .

We publish two versions: with breakpoints from all coins (the original) and from coins ≥ $100M. On CoinMarketCap, our CMKT correlates 0.95 with the published series over the sample of the original paper (2014 to July 2020) and 0.99 from 2018. Size and momentum agree less closely (0.47–0.80), because our cleaning is stricter and the data are a later vintage.

The specification multiverse

The standard was selected from 249,024 factor sets. They combine all ten design choices of Fieberg et al. (2024) with their options, except that we use two of their four data sources (they also use CoinPaprika and CryptoCompare):

Design choice Options
Data source CoinMarketCap, CoinGecko
Winsorisation (both tails, weekly) none, 0.1%, 0.5%, 1%
Stablecoins in, out
Minimum age 0, 26, 52 weeks
Minimum market cap none, $1M
Minimum price none, $1
Portfolios 3, 5, 10
Minimum coins per extreme portfolio 1, 5, 10
Weighting equal, value, capped value (80th percentile)
Implementation lag 0, 1 day

We add three choices of our own:

  • breakpoint universe: all coins, top 100, or coins ≥ $100M;
  • week calendar: 8 options;
  • the exact Liu et al. construction, as 24 additional factor sets per calendar.

The weekly factors of every world, their Sharpe ratios, the non-standard errors and the complete evaluation are in the research release on Hugging Face.

Pipeline

  1. Daily: CoinMarketCap listings (all coins) and a CoinGecko market snapshot (all coins) are stored, so that coins which later die are already archived. A validation job then runs content checks: coins per day, Bitcoin present, prices present, agreement between the two sources, and crypto2 against the official APIs.
  2. Monday:
    • the coin register is updated for new coins;
    • the daily and weekly panels are rebuilt and the standard factors computed;
    • the validation gate checks freshness, missing values, plausibility, the number of coins and the agreement between sources;
    • only if every check passes are the files published here and on Hugging Face.

References

Fama, E. F., & French, K. R. (1993). Common risk factors in the returns on stocks and bonds. Journal of Financial Economics, 33(1), 3–56.
Fieberg, C., Günther, S., Poddig, T., & Zaremba, A. (2024). Non-standard errors in the cryptocurrency world. International Review of Financial Analysis, 92, 103106.
Fieberg, C., Liedtke, G., Poddig, T., Walker, T., & Zaremba, A. (2025). A trend factor for the cross-section of cryptocurrency returns. Journal of Financial and Quantitative Analysis, 60(7), 3116–3153.
Ince, O. S., & Porter, R. B. (2006). Individual equity return data from Thomson Datastream: Handle with care! Journal of Financial Research, 29(4), 463–479.
Liu, Y., Tsyvinski, A., & Wu, X. (2022). Common risk factors in cryptocurrency. The Journal of Finance, 77(2), 1133–1177. https://doi.org/10.1111/jofi.13119