This paper examines whether recent cryptocurrency returns contain useful linear information about subsequent daily returns, and whether the remaining forecast errors exhibit economically important risk structure. The sample contains 2,191 aligned daily simple returns for Bitcoin (BTC), BNB and Cardano (ADA), covering 02 May 2020 to 01 May 2026.
Autoregressive models from AR(0) to AR(20) are estimated by maximum likelihood. Model selection balances parsimony, stationarity and residual whiteness. The resulting specifications are BTC AR(9), BNB AR(18) and ADA AR(9). The models remove most short-to-medium-lag linear autocorrelation, but the residual analysis shows that conditional-mean predictability is small relative to risk. Residuals are heavy-tailed, non-normal and volatility-clustered, while cross-asset correlations limit the diversification available from holding the three cryptocurrencies together.
Central finding. The defining statistical feature is not strong predictability in the mean. It is large, time-varying and heavy-tailed uncertainty around a comparatively weak autoregressive signal.
The analysis uses daily closing prices for BTC, BNB and ADA. The allocated price window is 1 May 2020 to 1 May 2026, inclusive. Raw histories extend beyond this window so that coverage can be verified before filtering.
For asset \(j\), the simple daily return is
\[ r_{j,t}=\frac{P_{j,t}}{P_{j,t-1}}-1, \]
where \(P_{j,t}\) is the closing price on day \(t\). The first return is dated 2 May 2020 because the 1 May closing price is required as its denominator. Cryptocurrency markets operate every day, so the expected sequence contains calendar days rather than business days.
| Asset | Raw rows | Raw start | Raw end | Missing closes | Duplicate dates | Retained prices | Daily returns |
|---|---|---|---|---|---|---|---|
| BTC | 2398 | 2020-01-10 | 2026-08-03 | 0 | 0 | 2192 | 2191 |
| BNB | 2398 | 2020-01-10 | 2026-08-03 | 0 | 0 | 2192 | 2191 |
| ADA | 2398 | 2020-01-10 | 2026-08-03 | 0 | 0 | 2192 | 2191 |
The raw histories contain no missing closing prices or duplicate dates. Filtering leaves 2,192 prices and 2,191 daily returns for each asset, with complete alignment across the three series.
| Asset | Mean return | Daily SD | Minimum | Median | Maximum | Skewness | Excess kurtosis | Annualised volatility |
|---|---|---|---|---|---|---|---|---|
| BTC | 0.145% | 3.026% | -15.975% | 0.041% | 18.746% | 0.121 | 3.742 | 57.8% |
| BNB | 0.249% | 4.263% | -33.266% | 0.130% | 69.760% | 2.438 | 40.348 | 81.4% |
| ADA | 0.201% | 5.178% | -26.009% | -0.080% | 71.328% | 1.732 | 19.399 | 98.9% |
Daily simple returns. Volatility is visibly clustered rather than constant through time.
For each asset, the conditional mean is represented by an AR(\(p\)) process:
\[ r_t-\mu=\phi_1(r_{t-1}-\mu)+\cdots+\phi_p(r_{t-p}-\mu)+\varepsilon_t, \]
where \(\mu\) is the unconditional
mean, \(\phi_i\) measures linear
dependence at lag \(i\), and \(\varepsilon_t\) is the innovation or
one-step prediction error. R’s arima() function uses this
centred-mean parameterisation when include.mean = TRUE for
an undifferenced ARMA model.
Candidate orders \(p=0,\ldots,20\) are estimated using maximum likelihood. AR(0) is included as a white-noise benchmark. A model enters the final candidate set when:
Among eligible candidates, the model with the smallest Bayesian Information Criterion (BIC) is selected. BIC penalises unnecessary parameters more strongly as sample size increases, while the residual condition prevents selection of a parsimonious model that leaves material linear dependence unmodelled.
BIC by AR order. Coloured points satisfy the stationarity and residual-whiteness criteria; red points are selected.
| Asset | Model | BIC | Ljung-Box p(20) | Ljung-Box p(30) | Minimum root modulus |
|---|---|---|---|---|---|
| BTC | AR(9) | -9,044.92 | 0.770 | 0.111 | 1.290 |
| BNB | AR(18) | -7,537.21 | 0.634 | 0.121 | 1.108 |
| ADA | AR(9) | -6,692.04 | 0.193 | 0.162 | 1.297 |
All minimum root moduli exceed one, supporting stationarity. The lag-20 and lag-30 Ljung–Box p-values exceed 0.05, so the selected models remove the principal short-to-medium-lag linear dependence under the stated diagnostic rule.
The fitted orders should not be confused with economic importance. A higher order means that more lag terms were required to whiten the residuals; it does not mean that returns are strongly predictable. Most estimated coefficients are small.
| Asset | Lag | Estimate | Standard error | p-value |
|---|---|---|---|---|
| BTC | 8 | -0.0473 | 0.0213 | 0.027 |
| BTC | 9 | 0.0567 | 0.0214 | 0.008 |
| BNB | 1 | -0.0452 | 0.0213 | 0.034 |
| BNB | 2 | 0.0772 | 0.0213 | <0.001 |
| BNB | 9 | 0.0503 | 0.0212 | 0.018 |
| BNB | 10 | 0.1276 | 0.0212 | <0.001 |
| BNB | 13 | 0.0450 | 0.0214 | 0.035 |
| BNB | 18 | 0.0659 | 0.0213 | 0.002 |
| ADA | 9 | 0.0643 | 0.0213 | 0.003 |
BTC’s clearest effects occur at lags 8 and 9, with a negative lag-8 coefficient followed by a positive lag-9 coefficient. BNB displays the broadest lag structure; its largest coefficient occurs at lag 10. ADA’s only 5%-significant AR coefficient is positive at lag 9. These estimates are consistent with a mixture of weak short-run reversal and lagged continuation, rather than a simple universal momentum rule.
| Asset | Model | N | Mean | Daily SD | Minimum | Q1 | Median | Q3 | Maximum | Skewness | Excess kurtosis | Annualised volatility |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| BTC | AR(9) | 2,191 | 0.00000% | 3.013% | -16.219% | -1.430% | -0.076% | 1.306% | 18.254% | 0.107 | 3.667 | 57.6% |
| BNB | AR(18) | 2,191 | -0.00005% | 4.184% | -32.151% | -1.709% | -0.098% | 1.621% | 61.627% | 1.774 | 29.257 | 79.9% |
| ADA | AR(9) | 2,191 | 0.00002% | 5.155% | -25.234% | -2.688% | -0.313% | 2.127% | 71.908% | 1.797 | 20.203 | 98.5% |
Residual means are effectively zero, as expected for fitted models with a mean term. The residual standard-deviation ranking is unambiguous: BTC is lowest, BNB is intermediate and ADA is highest. The AR components reduce dispersion only modestly relative to the raw returns, indicating that the majority of daily variation remains innovation rather than forecastable conditional mean.
BNB and ADA show pronounced positive skewness because of rare, exceptionally large positive residuals. This should not be read as an absence of downside risk: their minimum residuals are also economically large. Excess kurtosis is positive for every asset and extreme for BNB and ADA, implying far more tail mass than a Gaussian model would assign.
Residual histograms and normal Q<U+2013>Q plots. Systematic departures from the reference lines indicate non-normal tails.
| Asset | LB p(10) | LB p(20) | LB p(30) | LB p(50) | Jarque-Bera p | ARCH-LM(10) p |
|---|---|---|---|---|---|---|
| BTC | 0.530 | 0.770 | 0.111 | 0.007 | <0.001 | <0.001 |
| BNB | 0.824 | 0.634 | 0.121 | 0.108 | <0.001 | <0.001 |
| ADA | 0.106 | 0.193 | 0.162 | 0.136 | <0.001 | <0.001 |
The Jarque–Bera tests reject normality for all three residual series. The ARCH-LM tests also reject constant conditional variance, showing that large squared residuals are serially related. In economic terms, risk arrives in clusters rather than at a stable daily rate.
BTC is the only asset that fails the lag-50 Ljung–Box check, with (p=0.007). Its AR(9) model is adequate under the lag-20 and lag-30 selection rule but leaves some longer-horizon linear structure. This is a reported limitation rather than a reason to keep adding lags without an out-of-sample test.
Thirty-day rolling annualised residual volatility. A value of 1 represents 100% annualised volatility.
| BTC | BNB | ADA | |
|---|---|---|---|
| BTC | 1.000 | 0.631 | 0.645 |
| BNB | 0.631 | 1.000 | 0.550 |
| ADA | 0.645 | 0.550 | 1.000 |
All pairwise correlations are positive and moderately large. BTC–ADA is the strongest pair at 0.645, followed by BTC–BNB at 0.631. BNB–ADA is lower but remains positive at 0.550. Holding all three assets therefore provides some idiosyncratic diversification, but it does not remove the shared cryptocurrency-market component.
The fitted AR equations can produce one-day conditional-mean forecasts. The coefficient magnitudes and residual dispersion show, however, that the predictable component is small relative to the innovation. AR forecasts are better interpreted as a minor timing or position-sizing overlay than as a stand-alone directional strategy.
A practical signal should include a no-trade region. A forecast should trigger a position change only when its expected benefit is large enough to exceed estimated fees, bid–ask spread and slippage. Otherwise, statistical predictability can be converted into negative economic performance through turnover.
Fixed dollar positions would generate very different risk contributions across the three assets. For an equal risk budget, ADA would ordinarily receive the smallest notional position, BNB an intermediate position and BTC the largest. Because volatility changes markedly through time, these weights should respond to a forecast or recent estimate of conditional volatility rather than rely only on full-sample standard deviations.
The significant ARCH effects provide a direct statistical motivation for extending the mean equation to an AR–GARCH specification. A heavy-tailed innovation distribution, such as Student-\(t\), would address both volatility clustering and the observed departure from Gaussian tails.
Gaussian risk calculations would understate the probability of extreme residuals, especially for BNB and ADA. Historical simulation, stress testing and Expected Shortfall are more defensible starting points than relying only on a normal Value-at-Risk calculation. Position limits and drawdown controls remain necessary because a favourable average or positively skewed distribution does not rule out severe losses.
Portfolio weights should use the covariance matrix rather than equal dollar allocations. The assets differ substantially in residual volatility and are positively correlated. A covariance-aware or risk-budgeting approach would explicitly account for both facts. Broader diversification would require assets outside this three-coin set because common cryptocurrency-market movements remain material.
No strategy should be assessed using the same sample that selected the model. A defensible evaluation would:
The results describe in-sample linear dependence in one historical window. They do not establish stable causal relationships or guaranteed trading profits. Cryptocurrency market structure can change, making fixed coefficients unstable. The analysis also focuses on closing-price returns and does not include volume, liquidity, order-book conditions, macroeconomic variables or cross-asset lead–lag information.
The AR model addresses the conditional mean but assumes a simple innovation process during estimation. The diagnostics show that this is incomplete: variance is time-varying and tails are non-normal. Finally, statistical significance is not equivalent to economic significance. The small coefficients must be evaluated against implementation costs and forecast uncertainty.
BTC AR(9), BNB AR(18) and ADA AR(9) provide adequate in-sample descriptions of short-to-medium-lag linear dependence under the adopted BIC and residual-whiteness rule. The models are stationary and their residual means are close to zero. Nevertheless, the conditional-mean signal is weak relative to residual risk.
The most robust empirical findings concern the distribution and timing of risk. ADA has the highest residual volatility, followed by BNB and BTC. BNB and ADA are highly right-skewed and heavy-tailed; all three residual series reject normality and display volatility clustering. Positive cross-asset correlations further limit diversification within the selected cryptocurrency set.
Accordingly, the AR forecasts are most defensible as conservative tactical inputs embedded within a wider risk system. That system should use volatility-responsive sizing, heavy-tail risk measures, covariance-aware allocation, explicit cost thresholds and strict out-of-sample validation. An AR–GARCH model with heavy-tailed innovations is a natural next modelling step.
stats
documentation. https://stat.ethz.ch/R-manual/R-devel/library/stats/html/arima.html| Asset | AR order | Term | Estimate | Standard error | z statistic | p-value | Significant at 5% |
|---|---|---|---|---|---|---|---|
| BTC | 9 | ar1 | -0.034989 | 0.021328 | -1.641 | 0.101 | FALSE |
| BTC | 9 | ar2 | 0.020288 | 0.021312 | 0.952 | 0.341 | FALSE |
| BTC | 9 | ar3 | -0.001696 | 0.021313 | -0.080 | 0.937 | FALSE |
| BTC | 9 | ar4 | 0.017802 | 0.021311 | 0.835 | 0.404 | FALSE |
| BTC | 9 | ar5 | 0.020179 | 0.021311 | 0.947 | 0.344 | FALSE |
| BTC | 9 | ar6 | 0.004563 | 0.021336 | 0.214 | 0.831 | FALSE |
| BTC | 9 | ar7 | 0.001961 | 0.021332 | 0.092 | 0.927 | FALSE |
| BTC | 9 | ar8 | -0.047296 | 0.021330 | -2.217 | 0.027 | TRUE |
| BTC | 9 | ar9 | 0.056742 | 0.021375 | 2.655 | 0.008 | TRUE |
| BTC | 9 | intercept | 0.001451 | 0.000669 | 2.169 | 0.030 | TRUE |
| BNB | 18 | ar1 | -0.045161 | 0.021312 | -2.119 | 0.034 | TRUE |
| BNB | 18 | ar2 | 0.077220 | 0.021327 | 3.621 | <0.001 | TRUE |
| BNB | 18 | ar3 | -0.011745 | 0.021385 | -0.549 | 0.583 | FALSE |
| BNB | 18 | ar4 | 0.025955 | 0.021381 | 1.214 | 0.225 | FALSE |
| BNB | 18 | ar5 | -0.015702 | 0.021377 | -0.735 | 0.463 | FALSE |
| BNB | 18 | ar6 | 0.009196 | 0.021361 | 0.431 | 0.667 | FALSE |
| BNB | 18 | ar7 | 0.027639 | 0.021352 | 1.294 | 0.196 | FALSE |
| BNB | 18 | ar8 | -0.002015 | 0.021356 | -0.094 | 0.925 | FALSE |
| BNB | 18 | ar9 | 0.050256 | 0.021200 | 2.371 | 0.018 | TRUE |
| BNB | 18 | ar10 | 0.127634 | 0.021199 | 6.021 | <0.001 | TRUE |
| BNB | 18 | ar11 | -0.001988 | 0.021380 | -0.093 | 0.926 | FALSE |
| BNB | 18 | ar12 | -0.021564 | 0.021371 | -1.009 | 0.313 | FALSE |
| BNB | 18 | ar13 | 0.044953 | 0.021371 | 2.103 | 0.035 | TRUE |
| BNB | 18 | ar14 | -0.016392 | 0.021393 | -0.766 | 0.444 | FALSE |
| BNB | 18 | ar15 | -0.006419 | 0.021384 | -0.300 | 0.764 | FALSE |
| BNB | 18 | ar16 | 0.016783 | 0.021380 | 0.785 | 0.432 | FALSE |
| BNB | 18 | ar17 | -0.015626 | 0.021314 | -0.733 | 0.463 | FALSE |
| BNB | 18 | ar18 | 0.065895 | 0.021290 | 3.095 | 0.002 | TRUE |
| BNB | 18 | intercept | 0.002489 | 0.001291 | 1.928 | 0.054 | FALSE |
| ADA | 9 | ar1 | -0.037010 | 0.021315 | -1.736 | 0.083 | FALSE |
| ADA | 9 | ar2 | 0.031125 | 0.021320 | 1.460 | 0.144 | FALSE |
| ADA | 9 | ar3 | 0.008014 | 0.021322 | 0.376 | 0.707 | FALSE |
| ADA | 9 | ar4 | 0.013714 | 0.021318 | 0.643 | 0.520 | FALSE |
| ADA | 9 | ar5 | -0.025456 | 0.021309 | -1.195 | 0.232 | FALSE |
| ADA | 9 | ar6 | 0.007071 | 0.021310 | 0.332 | 0.740 | FALSE |
| ADA | 9 | ar7 | 0.021158 | 0.021306 | 0.993 | 0.321 | FALSE |
| ADA | 9 | ar8 | -0.024494 | 0.021297 | -1.150 | 0.250 | FALSE |
| ADA | 9 | ar9 | 0.064275 | 0.021293 | 3.019 | 0.003 | TRUE |
| ADA | 9 | intercept | 0.002006 | 0.001169 | 1.715 | 0.086 | FALSE |
The R Markdown document expects the following files in the same project:
Assignment1_Nathan_Foale/
├── Crypto_AR_Modelling_Paper.Rmd
├── Question1_AR_Analysis.R
└── raw/
├── BTC_raw.csv
├── BNB_raw.csv
└── ADA_raw.csv
Open Crypto_AR_Modelling_Paper.Rmd in RStudio and select
Knit. The analysis script uses only base R;
rmarkdown and knitr are required to produce
the HTML document.
## R version 4.4.3 (2025-02-28)
## Platform: aarch64-apple-darwin24.2.0
## Running under: macOS Sequoia 15.2
##
## Matrix products: default
## BLAS: /opt/homebrew/Cellar/openblas/0.3.29/lib/libopenblasp-r0.3.29.dylib
## LAPACK: /opt/homebrew/Cellar/r/4.4.3_1/lib/R/lib/libRlapack.dylib; LAPACK version 3.12.0
##
## locale:
## [1] C.UTF-8/C/C.UTF-8/C/C.UTF-8/C.UTF-8
##
## time zone: Australia/Melbourne
## tzcode source: internal
##
## attached base packages:
## [1] stats graphics grDevices utils datasets methods base
##
## loaded via a namespace (and not attached):
## [1] digest_0.6.39 R6_2.6.1 fastmap_1.2.0 xfun_0.57
## [5] cachem_1.1.0 knitr_1.51 htmltools_0.5.9 rmarkdown_2.30
## [9] lifecycle_1.0.5 cli_3.6.5 sass_0.4.10 jquerylib_0.1.4
## [13] compiler_4.4.3 tools_4.4.3 evaluate_1.0.5 bslib_0.10.0
## [17] yaml_2.3.12 rlang_1.1.7 jsonlite_2.0.0