High-Frequency Index Analysis

• High-Frequency Index is an index observed at significantly finer time scales than traditional aggregates, capturing intraday market and economic fluctuations.
• These indices exhibit distinctive statistical regularities such as multifractality, heavy-tailed distributions, scaling symmetries, and long memory that challenge conventional models.
• Applications range from real-time volatility forecasting in financial markets to constructing pseudo-weekly economic indicators through dynamic imputation and adaptive learning.

Searching arXiv for the cited papers and related work on high-frequency financial and economic indices. arXiv search query: "(Suárez-García et al., 2013) Multifractality and long memory of a financial index" arXiv search query: "(Suárez-García et al., 2012) Scaling, stability and distribution of the high-frequency returns of the IBEX35 index" arXiv search query: "(Ng et al., 2023) Constructing High Frequency Economic Indicators by Imputation" arXiv search query: "(García-Albán et al., 10 Jul 2025) Tracking the economy at high frequency" A high-frequency index is an index observed or constructed on a time scale substantially finer than conventional daily, monthly, or quarterly reporting. In the literature represented here, the term encompasses intraday financial indices such as the IBEX35, CSI 300 Index Futures, and the VIX observed at 15-second, 5-minute, or 10-minute resolution, as well as synthetic economic indices built at weekly or pseudo-weekly frequency from mixed-frequency official statistics. Across these settings, high-frequency indices are studied not merely as faster versions of low-frequency aggregates, but as objects with distinctive scaling laws, heavy-tailed distributions, long-memory or serial-correlation structures, intraday periodicities, and in some cases multifractal organization (Suárez-García et al., 2013, Ng et al., 2023, García-Albán et al., 10 Jul 2025).

1. Scope, measurement, and representative forms

High-frequency indices arise in two principal forms. The first is the directly observed market index or index-linked price process sampled intraday. Examples in the cited literature include the IBEX35 at 15-second intervals, CSI 300 Index Futures summarized in 5-minute brackets from order-book data, the VIX at 15-second tick resolution, and Bitcoin market index data at 10-minute intervals (Suárez-García et al., 2013, Yang, 2021, Shang et al., 2018, Tang et al., 2024). The second is the constructed macroeconomic indicator, where higher-frequency values are treated as missing and imputed, or where weekly, monthly, and quarterly indicators are harmonized into a pseudo-weekly latent factor (Ng et al., 2023, García-Albán et al., 10 Jul 2025).

Domain	Example	Frequency
Financial market index	IBEX35 returns	15 seconds
Order-book index futures	CSI 300 Index Futures	5 minutes
Volatility index	VIX	15 seconds
Crypto market index	Bitcoin market index	10 minutes
Economic activity index	HFEI	pseudo-weekly

The measurement architecture depends on the application. For intraday finance, the object is commonly a return sequence, such as $r_t^{(l)} = \ln s_t^{(l)} - \ln s_{t-1}^{(l)}$, or a cumulative intraday return curve, such as $$R_i(t_j) = 100 \times [\ln P_i(t_j) - \ln P_i(t_1)]$$ (Baldovin et al., 2012, Shang et al., 2018). In macroeconomic applications, the target is often a low-frequency diffusion index used as an anchor, while the desired higher-frequency values are treated as missing and inferred from higher-frequency covariates (Ng et al., 2023). A further refinement is the pseudo-week structure, in which each month is divided into four fixed periods, maintaining temporal consistency and avoiding irregular aggregation artifacts (García-Albán et al., 10 Jul 2025).

2. Core statistical regularities

The empirical literature consistently associates high-frequency financial indices with several stylized facts. For the IBEX35, the reported findings include absence of linear autocorrelations in raw returns, long-memory and volatility clustering in absolute or squared returns, and non-Gaussian leptokurtic distributions with scaling symmetry (Suárez-García et al., 2013). The same study links these properties to a wide singularity spectrum and to temporal dependence in nonlinear moments rather than to raw-return autocorrelation (Suárez-García et al., 2013).

Distributional non-Gaussianity appears in several forms. For IBEX35 high-frequency returns, the return distribution appears approximately stable across time scales after square-root-of-time rescaling, but maximum likelihood estimation and goodness-of-fit tests reject the Lévy-stable law as a plausible underlying probabilistic model (Suárez-García et al., 2012). For Bitcoin intraday returns, heavy tails are described using a $q$-Gaussian distribution, with fitted values $q = 1.51 \pm 0.02$ in Period 1 and $q = 1.50 \pm 0.02$ in Period 2 from tail-based estimation (Tang et al., 2024). In both periods, the Bitcoin market index exhibits an anomalous diffusion process, moving from subdiffusion at short horizons to weak superdiffusion at longer horizons (Tang et al., 2024).

These regularities coexist with sharp time-scale asymmetries. The ensemble autocorrelation of Bitcoin returns decays rapidly, whereas the autocorrelation of absolute returns follows a power law (Tang et al., 2024). For IBEX35, raw returns are serially uncorrelated, but absolute returns show persistent long-range correlations (Suárez-García et al., 2013). A common implication is that the absence of linear predictability in returns does not preclude pronounced dependence in volatility, magnitude, or other nonlinear functionals.

3. Multifractality, singularity spectra, and long memory

The most explicit multifractal treatment in the cited corpus is the analysis of IBEX35 high-frequency returns by Multifractal Detrended Fluctuation Analysis (MFDFA). The local Hölder exponent is defined through

$$|X(t+\epsilon)-X(t)| \sim C(t)\epsilon^{\alpha(t)}, \qquad \epsilon \rightarrow 0,$$

with monofractals characterized by constant $\alpha(t)$ and multifractals by varying $\alpha(t)$ (Suárez-García et al., 2013). Global scaling is expressed as

$\mathbb{E}[|X(t)|^q] \sim c(q) t^{\tau(q)+1},$

and for finite series via the partition function

$$Z_q(\epsilon) = \sum_{n=0}^{N-1} |X(n\epsilon+\epsilon) - X(n\epsilon)|^q, \qquad \mathbb{E}[Z_q(\epsilon)] \sim \epsilon^{\tau(q)}.$$

The singularity spectrum follows by Legendre transform,

$$R_i(t_j) = 100 \times [\ln P_i(t_j) - \ln P_i(t_1)]$$0

Its width quantifies the degree of multifractality (Suárez-García et al., 2013).

In MFDFA, the sample is divided into $$R_i(t_j) = 100 \times [\ln P_i(t_j) - \ln P_i(t_1)]$$1 non-overlapping segments of length $$R_i(t_j) = 100 \times [\ln P_i(t_j) - \ln P_i(t_1)]$$2, each segment is detrended by fitting and subtracting a polynomial of order $$R_i(t_j) = 100 \times [\ln P_i(t_j) - \ln P_i(t_1)]$$3, and the residuals are aggregated into a $$R_i(t_j) = 100 \times [\ln P_i(t_j) - \ln P_i(t_1)]$$4-order fluctuation function satisfying

$$R_i(t_j) = 100 \times [\ln P_i(t_j) - \ln P_i(t_1)]$$5

The relation

$$R_i(t_j) = 100 \times [\ln P_i(t_j) - \ln P_i(t_1)]$$6

connects the fluctuation exponent to multifractal scaling (Suárez-García et al., 2013).

Applied to the IBEX35 over 2009–2010 at 15-second sampling, this procedure yields a strongly nonlinear $$R_i(t_j) = 100 \times [\ln P_i(t_j) - \ln P_i(t_1)]$$7 and a wide singularity spectrum with $$R_i(t_j) = 100 \times [\ln P_i(t_j) - \ln P_i(t_1)]$$8 and $$R_i(t_j) = 100 \times [\ln P_i(t_j) - \ln P_i(t_1)]$$9, indicating intermittent and heterogeneous fluctuations (Suárez-García et al., 2013). The same study concludes that the multifractality primarily originates from long-range temporal correlations in nonlinear moments such as volatility rather than from the fat tails of the return distribution (Suárez-García et al., 2013).

The long-memory structure is decomposed into two components. One is a high-frequency component linked to daily cycles of market activity and information arrival, including the “lunch effect.” The other is a slowly-varying component that persists over long periods and shows no apparent relation with human economic cycles; it is therefore postulated to be endogenous to market dynamics (Suárez-García et al., 2013). This suggests that multifractal structure in a high-frequency index need not be reducible to scheduled news arrival or purely exogenous shocks.

4. Distributional modeling, scaling, and information flow

A central issue in the analysis of high-frequency indices is whether observed scaling symmetry implies genuine stability in the sense of Lévy aggregation. For IBEX35 high-frequency returns, the empirical distribution appears approximately stable from 15 seconds up to several hours when rescaled with $q$0, and the exponent $q$1 is derived via Detrended Fluctuation Analysis, consistent with i.i.d. variables of finite variance rather than with stable laws requiring exponent $q$2 (Suárez-García et al., 2012). Reshuffling the data destroys this apparent stability and induces fast Gaussian convergence, supporting the interpretation that the observed scaling is due to temporal correlations or non-stationarities rather than to an underlying Lévy-stable law (Suárez-García et al., 2012).

Within parametric families, the same IBEX35 study evaluates Lévy-stable laws and subclasses of the Generalized Hyperbolic family using maximum likelihood estimation together with the $q$3, Kolmogorov–Smirnov, and Anderson–Darling tests (Suárez-García et al., 2012). No tested distribution formally passes all goodness-of-fit tests at standard significance levels, given the immense sample size, but the Normal Inverse Gaussian distribution provides a better overall fit than Lévy-stable laws or the skewed Student’s $q$4 (Suárez-García et al., 2012). Tail estimation gives a right-tail exponent $q$5 and a left-tail exponent $q$6, values that lie well above the stability region $q$7 (Suárez-García et al., 2012).

High-frequency index analysis also includes lead-lag and feedback structure across related contracts. In the Chinese stock index futures market, the most liquid near-month contract consistently leads longer-dated contracts by one tick across all indices and most days, and the lead-lag spread exhibits a negative feedback effect on the leading asset that can predict returns of the leader (Li et al., 6 Jan 2025). Liquidity is identified as the main driver of the direction and strength of these relationships (Li et al., 6 Jan 2025). In the VIX complex during January and February 2018, the evidence reported in the abstract indicates bidirectional causality between VIX spot and the implied volatility of S&P 500 options, together with a significant mean-reverting equilibrium relationship in a vector autoregression error-correction framework (2206.13138).

These results complicate a common simplification according to which a high-frequency index is merely a noisy version of an efficient low-frequency process. The cited evidence instead points to microstructure-sensitive dependence, feedback, and cross-instrument price discovery that are visible only at granular time scales.

5. Forecasting, adaptive learning, and trading uses

Forecasting methods for high-frequency indices are correspondingly heterogeneous. For the VIX, one approach treats each day’s 15-second tick sequence as a curve and models it as a functional time series in $q$8, using decomposition

$q$9

Dynamic updating methods then refine one-day-ahead point and interval forecasts as partial intraday data arrive (Shang et al., 2018). The cited study reports that point and interval forecasts enjoy improved accuracy over conventional time series models, and that functional linear regression performs especially well in dynamic updating (Shang et al., 2018).

For CSI 300 Index Futures, a forecast-centric adaptive learning framework combines classical time-series models such as ARIMAX and VARMA with time-varying model selection based on recent forecast errors and penalties for model switching. The predictive target is written abstractly as

$q = 1.51 \pm 0.02$0

and the selected model is

$q = 1.51 \pm 0.02$1

The reported empirical result is that adaptive learning performs well compared with the top fixed models and can improve forecasting accuracy by being more stable and resilient to non-stationarity (Yang, 2021).

Scaling-based models have also been used for conditional density forecasting and intraday trading rules. For S&P 500 high-frequency data, the distribution of aggregated returns is represented by

$q = 1.51 \pm 0.02$2

with different empirical exponents in morning and afternoon sessions. A non-Markovian martingale model built from these scaling properties supports conditional density forecasts and quantile-based trend-following rules (Baldovin et al., 2012). The paper reports that the resulting strategy outperforms a benchmark asymmetric GARCH process in in-sample and out-of-sample tests and that observed profits stem from small linear correlations present in the data and absent in the martingale model (Baldovin et al., 2012).

At the same time, the trading literature emphasizes the fragility of apparent profitability. For stationary variants of MA, KDJ, and Bollinger-band rules applied to 15-second, 30-second, and 60-second CSI300 Index Futures data, several parameter combinations are significant when transaction costs are ignored, but once transaction costs are included, trading profits are eliminated completely (Chen et al., 2017). A related implication is that predictive structure in a high-frequency index can be statistically detectable without being economically exploitable after frictions.

Machine-learning augmentation appears in stochastic-volatility settings as well. For 1-minute CSI 300 index data, a generalized Barndorff-Nielsen and Shephard model introduces market information asynchrony, microstructure noise, and long-term dependence, while machine-learning and deep-learning algorithms are used to estimate a deterministic parameter governing future large jumps (Hui et al., 2022). The reported conclusion is that deterministic components of stochastic volatility processes can be captured over short and longer-term windows, with implications for jump prediction and realized-volatility-based monitoring (Hui et al., 2022).

6. High-frequency economic indices, aggregation, and interpretation

In macroeconomic applications, a high-frequency index is typically latent rather than directly quoted. One construction strategy begins with a low-frequency diffusion index that is already available and treats the unobserved high-frequency values as missing. These values are then imputed using multiple factors estimated from higher-frequency covariates (Ng et al., 2023). Static matrix completion assumes low rank but does not account for serial correlation in idiosyncratic errors, and the reported empirical finding is that such static procedures yield imprecise imputations irrespective of how the factors are estimated (Ng et al., 2023). By contrast, single-equation and systems-based dynamic procedures that model serial correlation produce imputed values closer to the observed low-frequency ones (Ng et al., 2023).

The dynamic single-equation form is

$q = 1.51 \pm 0.02$3

while the state-space factor form is

$q = 1.51 \pm 0.02$4

with Kalman filtering or smoothing used to impute missing high-frequency values (Ng et al., 2023). Empirical examples include a weekly version of the CFNAI and the reconstruction of monthly consumer sentiment before 1978, when the series was released only quarterly (Ng et al., 2023).

A more explicitly model-based construction is the High-Frequency Economic Index (HFEI), developed using a Bayesian Dynamic Factor Model with weekly, monthly, and quarterly official indicators, dynamic heterogeneity, and stochastic volatility (García-Albán et al., 10 Jul 2025). The observation equation is

$q = 1.51 \pm 0.02$5

and temporal consistency is enforced through a pseudo-week structure dividing each month into four fixed periods (García-Albán et al., 10 Jul 2025). Monthly and quarterly variables are linked to weekly latent variables through aggregation constraints such as

$q = 1.51 \pm 0.02$6

and

$q = 1.51 \pm 0.02$7

The resulting HFEI is interpreted as a pseudo-weekly coincident index of economic activity and is used to compute pseudo-weekly recession probabilities by a time-varying mean regime-switching model (García-Albán et al., 10 Jul 2025).

The macroeconomic literature therefore treats a high-frequency index not as a simple interpolation, but as a dynamically constrained latent process. A common misconception is that low-rank completion alone is sufficient; the cited results indicate instead that serial correlation and timing conventions are decisive for credible high-frequency reconstruction (Ng et al., 2023, García-Albán et al., 10 Jul 2025).