When Small Coins Roar: Rethinking Systemic Risk in Crypto Volatility Forecasting

TL;DR for operators

Crypto risk systems often watch the obvious giants. Bitcoin first. Ethereum second. Everything else somewhere in the dashboard’s lower intestine, where altcoins go to become colourful noise.

This paper argues that this is not enough.¹ In volatility forecasting, the asset that matters is not always the asset with the largest market capitalisation. It may be the asset whose volatility is transmitting stress into Bitcoin under the current market state. That is a different question, and it produces a different monitoring system.

The authors propose SA-Log-HAR, a state-adaptive extension of Log-HAR volatility forecasting. The model keeps the familiar multi-scale structure of HAR-style volatility models—daily, weekly, and monthly volatility components—but adds a spillover variable chosen from the crypto network. The key detail is that the spillover source is not fixed. It changes according to whether the target asset is in a low, normal, or high volatility state, and according to quantile-level spillover structure.

For a crypto risk desk, the operational takeaway is clear: do not merely ask, “How volatile is Bitcoin?” Ask, “Which asset is currently transmitting volatility into Bitcoin, and does that transmitter change in the tails?” The answer matters most when the market is already behaving badly, or unusually well, because the paper finds that spillovers amplify in both tails.

The paper’s empirical evidence is based on 5-minute Binance data for BTC, DASH, ETH, LTC, XLM, and XRP from 28 March 2019 to 30 March 2025. The final 300 days are reserved for out-of-sample forecasting in the main test, with additional robustness checks using a 500-day evaluation window, realised kernel volatility, and out-of-sample predictive performance tests.

This is not a magic volatility oracle. The study does not prove a universal causal law, and it does not hand over a live trading edge wrapped in academic ribbon. What it does offer is a useful design principle: crypto volatility forecasting should treat systemic influence as a dynamic network role, not as a synonym for size.

The useful mechanism is not “better HAR”; it is adaptive source selection

The paper’s contribution is easiest to understand if we begin with the mechanism rather than the model name.

A standard HAR-style realised volatility model tries to forecast future volatility using volatility measured over multiple horizons. The rough intuition is that different market participants operate on different clocks. Intraday traders, weekly rebalancers, and longer-horizon allocators all leave traces in the volatility process. A HAR model approximates that layered persistence using short-, medium-, and longer-window realised volatility components.

The Log-HAR version applies this logic to log-transformed realised volatility, which is useful in crypto because raw volatility can be violently non-normal. A polite spreadsheet might call this “heavy-tailed.” A trading desk would call it “Tuesday.”

The authors keep that multi-scale backbone. Their added idea is to inject information from the crypto market’s internal volatility network. In schematic form, the forecast becomes:

That second term is the novelty. The model does not simply include “ETH volatility” or “market volatility” as a fixed extra feature. It constructs a state-adaptive spillover variable.

The construction has three steps.

First, classify the target asset’s volatility state. For Bitcoin, the paper partitions the relevant volatility feature into low, normal, and high regimes using quantile thresholds. The paper discusses lower and upper quantiles such as 0.05 and 0.95, creating a left-tail, middle, and right-tail view of market conditions.

Second, identify the dominant spillover source for each state. The authors use a time-varying quantile spillover framework based on TVP-QVAR, Kalman filtering, and generalized forecast error variance decomposition. Put less elegantly but more usefully: they estimate who is transmitting volatility to whom, and how that relationship changes across time and quantiles.

Third, use the selected source’s lagged volatility as an explanatory variable in the Log-HAR model. So if XRP is the dominant transmitter to BTC in one regime, XRP’s relevant lagged volatility component enters the forecast. If LTC dominates in another, the source changes.

This is the important part. The paper is not only saying, “Add spillovers.” It is saying, “Add the spillover source that matters under the current volatility state.” Fixed features assume the same neighbour matters all the time. Crypto markets, being crypto markets, decline the invitation to be that tidy.

The model is therefore less about theoretical ornamentation and more about feature governance. It asks which external volatility signal should be allowed into the forecast, under which market condition, and why.

That is the part operators should care about.

Tail spillovers are not just larger; they are differently informative

The first empirical result is about the shape of spillovers across quantiles.

The paper reports that total volatility spillovers are relatively stable around the median, but strengthen in both tails. This is not surprising in the lazy sense—markets often become more connected under stress—but the two-tail result matters. It means the network does not only tighten during downside panic. It also becomes more connected during extreme upside or high-volatility states.

That matters for crypto because upside mania can be as operationally dangerous as downside collapse. Leverage expands. Funding stress appears in different places. Market makers widen spreads. Retail flows chase volatility. The exchange risk engine does not get to relax merely because the candles are green.

The paper uses multiple volatility features, including realised volatility, continuous and jump components, realised semivariance, and realised extreme-type measures. The appendix figures extend the total and net spillover analysis across these feature choices and quantiles. Their likely purpose is robustness and mechanism diagnosis: the authors are not trying to make a second thesis out of every appendix panel; they are checking whether the quantile-spillover pattern is visible across volatility representations.

The business translation is simple: a dashboard that only monitors average connectedness is under-instrumented. Median-market spillovers are the calm office-hours version of the system. Tail spillovers are the version that shows up during the incident call.

The market-cap shortcut breaks where risk transmission begins

The most business-relevant misconception is also the easiest one to hold: Ethereum is huge, therefore Ethereum must be the most important non-Bitcoin source of systemic volatility pressure.

The paper complicates that. In the authors’ BTC-centred spillover analysis, ETH often behaves more like a net receiver, while smaller-cap assets such as XRP and LTC emerge as stronger net transmitters in parts of the network. The paper’s language is careful: systemic influence is shaped more by functional centrality than by market size.

That distinction is worth spelling out.

Market capitalisation measures scale. Spillover centrality measures transmission role. Those two can overlap, but they are not the same object. A large asset can absorb shocks without being the main transmitter into Bitcoin. A smaller asset can occupy a more influential network position under particular volatility regimes.

For risk management, this changes the watchlist logic.

This is where the paper becomes useful beyond academic volatility modelling. It suggests a practical design principle for crypto risk systems: maintain separate rankings for capital importance, liquidity importance, and spillover importance. Combining them into one “top assets” list is administratively convenient and analytically sloppy. Naturally, dashboards love doing exactly that.

The forecast tests reward adaptation, but not uniformly

The authors test several model families: baseline Log-HAR variants, SA-Log-HAR variants with the state-adaptive spillover variable, and Lasso-SA-Log-HAR variants with regularisation. They compare forecasting performance over one-step, five-step, and 22-step horizons.

The 300-day out-of-sample Model Confidence Set tests show a horizon-specific pattern:

This is more interesting than a generic “new model wins” story.

At short and medium horizons, adaptive spillover information appears valuable. That makes intuitive sense. Nearby volatility is still close enough for network transmission to matter. The market’s current structure has not yet decayed into tomorrow’s stale dashboard.

At longer horizons, simpler Log-HAR-RV and SA-Log-HAR-RV specifications remain competitive. The authors interpret this through the chaotic and multifractal nature of Bitcoin volatility. In plainer English: the farther ahead you forecast, the more fragile highly decomposed volatility features become. Jump components and signed variations may capture meaningful microstructure detail, but detail is not always stability.

This is a useful warning for model builders. More refined volatility decomposition is not automatically better. A model can become exquisitely sensitive to yesterday’s plumbing while being less useful for next month’s risk budget.

The robustness checks are not decoration; they test three specific failure modes

The paper’s robustness section changes three things: the forecast window, the volatility measure, and the evaluation metric.

That is exactly what it should do. Each change targets a different way the main result could be fragile.

The out-of-sample evaluation is especially blunt. HAR-type models report out-of-sample $R^2$ values around 0.77 in the 300-day window and around 0.74–0.75 in the 500-day window, while several GARCH-type models produce much weaker or even negative results. Within the HAR family, SA and Lasso-SA variants add incremental improvement, with SA-Log-HAR-RS and Lasso-SA-Log-HAR-RS achieving the highest reported values in the relevant table.

The Clark–West tests report statistically significant improvements for HAR-type models, with p-values rounded to 0.000 in the table. That does not mean “perfect certainty,” obviously. It means the reported test strongly rejects the relevant no-improvement benchmark at the displayed precision.

The robustness results do not make the paper invincible. They make the claim narrower and more credible: in this dataset, with these six Binance-traded assets, these volatility estimators, and these forecast horizons, state-adaptive spillover information improves or supports HAR-style volatility forecasting, especially in short and medium horizons.

That is enough. A paper does not need to become a religion to be useful.

What Cognaptus infers for business use

The paper directly shows three things.

First, quantile spillovers in the studied crypto network amplify in both tails.

Second, systemic volatility influence can decouple from market capitalisation. ETH’s size does not automatically make it the dominant BTC volatility transmitter; XRP and LTC can matter more in the network role the paper measures.

Third, adding state-adaptive spillover variables to Log-HAR-style volatility models improves forecasting performance in the tested setup, with the strongest benefits depending on horizon and volatility feature.

Cognaptus infers a practical architecture from those results:

1. Separate asset-ranking systems. Do not use one ranking for liquidity, market cap, volatility contribution, and systemic transmission. They answer different questions.

2. Use regime-specific spillover monitors. A single average connectedness score is too blunt. Maintain separate views for normal conditions and tail conditions.

3. Feed diagnostics into forecasts. Spillover analysis should not live only in research slides. The paper’s core design is valuable because it turns network information into model input.

4. Match model complexity to horizon. Short-horizon forecasting may justify adaptive spillover and richer components. Longer-horizon forecasting may reward simpler realised-volatility structure.

5. Validate on operating constraints. A better statistical forecast is not automatically a better trading system. Before deployment, it must survive latency, liquidity, exchange fragmentation, fees, slippage, position limits, funding costs, and governance review. Yes, the boring list is where models go to become businesses.

The business value is therefore not “predict Bitcoin volatility perfectly.” It is cheaper and more adaptive diagnosis: knowing which asset’s volatility should matter today, under this state, at this horizon.

A useful implementation sketch for risk teams

A practical team would not need to copy the paper mechanically to benefit from it. The better route is to convert the method into a modular monitoring and forecasting pipeline.

The key governance rule is that source switching must be explainable. If the model changes the dominant transmitter from LTC to XRP, operators need to know whether this reflects a genuine network shift, a data-quality issue, or a transient liquidity distortion.

This is where the paper’s mechanism-first design helps. It is easier to audit a model that says, “I selected this asset because it was the dominant net pairwise transmitter under this quantile state,” than a model that says, “The neural net felt a disturbance in the Force.” The latter may still work. It is just less fun during a risk committee meeting.

Boundaries: where the result should not be over-sold

The paper’s evidence is useful, but its boundary is sharp.

The dataset contains six cryptocurrencies: BTC, DASH, ETH, LTC, XLM, and XRP. That is enough to study a compact BTC-centred network, but not enough to represent the full crypto market. It excludes stablecoins, DeFi governance tokens, newer high-beta assets, memecoins, liquid staking tokens, and cross-chain ecosystem effects. Some of those categories may transmit volatility very differently.

The source is Binance 5-minute data. That is a reasonable high-frequency venue choice, but crypto liquidity is fragmented across exchanges, derivatives markets, and offshore venues. A production system would need to test whether the same transmitter rankings hold across spot, perpetual futures, options, and exchange-specific microstructure.

The paper focuses on volatility forecasting, not causal identification. A coin appearing as a net transmitter in the spillover framework does not mean it “causes” Bitcoin volatility in the legal, structural, or macroeconomic sense. It means that within the model’s connectedness framework, its volatility innovations help explain forecast-error variance in the network.

The model also forecasts volatility, not returns. This should be obvious. It will not be obvious to someone trying to turn every volatility paper into a directional trading pitch by lunchtime.

Finally, forecast improvement is statistical. The paper does not test a complete trading strategy with costs, slippage, leverage constraints, collateral requirements, or liquidation mechanics. For market makers and risk managers, that is fine: volatility forecasts have value even when they are not converted directly into alpha. For automated trading, it is only the beginning.

The real lesson is diagnostic humility

The best part of the paper is not that it adds another acronym to the volatility shelf. Finance has enough acronyms to tile a small airport.

The useful lesson is diagnostic humility. Crypto markets do not transmit risk according to the hierarchy shown on a market-cap website. They transmit it through changing networks, and those networks behave differently in the tails.

SA-Log-HAR is one way to operationalise that lesson. It keeps a familiar volatility-forecasting backbone, then lets the relevant spillover source change with the market state. That combination is why the paper deserves attention: not because it defeats every model in every horizon, but because it points toward risk systems that are less static, less size-obsessed, and more aware of regime-specific transmission.

For operators, the message is simple. Bitcoin may be the asset you are forecasting. Ethereum may be the asset you are watching. But on the day the network starts shouting, the signal may come from somewhere smaller.

Small coins roar. Good risk systems should know when to listen.

Cognaptus: Automate the Present, Incubate the Future.