The Paglia River, a major tributary of the Tiber, is known for sudden floods, the most recent in 2012 causing serious damage to the town of Orvieto. Forecasting such events requires understanding the long-term memory of river flow. Classic studies, starting with Harold Hurst's analysis of the Nile, have shown that rivers can 'remember' the past longer than a random walk would suggest. However, for small and medium-sized Mediterranean rivers, such detailed studies are scarce. The Paglia data, collected by the Umbria regional service, provide a rare opportunity to test these ideas on a long, high-resolution series.
The authors used 377,399 streamflow observations (m³/s) from the Ponte dell'Adunata hydrometric station near Orvieto, covering the period from November 2, 1992, to May 13, 2014, at 30-minute intervals. Gaps (seven breaks longer than two days) were filled by linear interpolation in logarithms. To assess long-range dependence, they applied detrended fluctuation analysis (DFA) on sliding windows of 90, 180, and 400 days, shifted by one day. This yielded 7,773 estimates of the Hurst exponent for the 90-day windows. Additionally, they checked the data against Benford's law for the first two digits and approximated the rank distribution of streamflows with a discrete generalized beta distribution.
The average Hurst exponent across all windows was about 1/3, indicating pronounced anti-persistence: after high flow, the system tends to return to normal rather than continue rising. The estimates are stable: for 90-day windows, 90% of changes between adjacent windows lie in the range from -0.043 to 0.046. Spectral analysis of the Hurst exponent series revealed periodicities around 12, 6, 3, and 2.5 days. Monthly flow maxima show an increase in recent years, consistent with general climate change. The flow distribution is well described by a discrete generalized beta distribution with R² ≈ 0.986, typical for complex systems. The Benford's law test showed a significant discrepancy: the data do not follow the expected distribution of first digits, likely due to rounding of water level measurements to the nearest centimeter.
This work aligns with hydrological studies of long-term memory in river series, but differs from neighboring studies in the scale of data and the use of sliding windows to estimate the Hurst exponent. Unlike single estimates over the entire series, this approach reveals how the system's memory changes over time and responds to extreme events, such as the 2012 flood. This is the first such detailed test of anti-persistence for a river of this size in central Italy. The observed Benford's law discrepancy raises an important methodological question about the quality of hydrological measurements, which is rarely discussed in the literature.
The topic could evolve toward more precise testing of the link between Paglia streamflow and large-scale climate oscillations, such as El Niño and the North Atlantic Oscillation. This would require comparisons with indices of these phenomena and possibly causal analysis. Also promising is the application of alternative Hurst exponent estimation methods that are robust to non-stationarity and heteroscedasticity, to check whether the observed periodicities are artifacts of DFA. Finally, improving water level measurement techniques (e.g., to millimeter precision) would allow testing whether the Benford anomaly disappears.
The results could be useful for hydrological forecasting in regions with similar climates, especially for flood risk assessment. The sliding window methodology is applicable to any long series of ecological or geophysical observations.
The immediate next step is to test the relationship between the Hurst exponent and El Niño and North Atlantic Oscillation indices over the same time intervals. It is also worth repeating the analysis with higher-resolution data if they become available.
In hydrology, the question remains open as to how long-term memory in river flow relates to the physical mechanisms of the catchment and atmospheric circulation. The anti-persistence of the Paglia indicates strong negative feedbacks, but their nature is unclear. Furthermore, the data quality issue revealed by Benford's law is relevant across the field: without reliable measurements, any statistical conclusions about long-range dependence may be distorted.
The two most frequent streamflow values in the dataset each occur more than 10,000 times. This results from water levels being measured to the nearest centimeter and streamflow being computed via a fixed rating curve, creating 'steps' in the data.