% ==============================================================================
% RegFDC-Cauca v1.0.0 — Scientific manuscript
% Format: engrXiv
% Language: English
% ==============================================================================
\documentclass[a4paper]{article}

%% Language and font encodings
\usepackage[T1]{fontenc}
\usepackage[utf8]{inputenc}
\usepackage[english]{babel}

\usepackage[colorlinks=true,
            allcolors=blue,
            pdftitle={RegFDC-Cauca v1.0.0: Flow Duration Curve Regionalization via the Index-Flow Method with Long-Memory in Ungauged Basins of the Cauca River System, Colombia},
            pdfauthor={Mauricio Javier Victoria Niño},
            pdfsubject={Computational hydrology; Flow duration curves; Regionalization; ARFIMA; Hurst; Colombian Andes},
            pdfkeywords={flow duration curves; regionalization; index-flow method; Ward D2; ARFIMA; Hurst exponent; Cauca River; ungauged basins}
            ]{hyperref}

\urlstyle{tt}
\newcommand{\email}[1]{\href{mailto:#1}{\tt{\nolinkurl{#1}}}}
\newcommand{\orcid}[1]{ORCID: \href{https://orcid.org/#1}{\tt{\nolinkurl{#1}}}}

\usepackage[sfdefault,lf]{carlito}
\usepackage[parfill]{parskip}
\renewcommand*\oldstylenums[1]{\carlitoOsF #1}
\usepackage{fancyhdr}
\usepackage{authblk}
\setlength{\headheight}{41pt}

%% Page size and margins
\usepackage[a4paper,top=3cm,bottom=2cm,left=3cm,right=3cm,marginparwidth=1.75cm]{geometry}

%% Useful packages
\usepackage{amsmath, amssymb, amsfonts}
\usepackage{graphicx}
\usepackage{booktabs}
\usepackage{array}
\usepackage{multirow}
\usepackage{longtable}
\usepackage{caption}
\usepackage{subcaption}
\usepackage{float}
\usepackage{siunitx}
\usepackage{doi}
\usepackage{setspace}

\sisetup{
    per-mode          = symbol,
    inter-unit-product = \ensuremath{{}\cdot{}},
    output-decimal-marker = {.}
}

%% 1.5 line spacing
\onehalfspacing

%% engrXiv-style header
\pagestyle{fancy}
\fancyhf{}
\fancyhead[L]{Posted: \today}
\fancyhead[R]{\includegraphics[width=4cm]{engrXiv_banner.png}}
\fancyfoot[C]{\thepage}

%% Title and author
\title{\textbf{RegFDC-Cauca v1.0.0: Flow Duration Curve Regionalization via
the Index-Flow Method with Long-Memory in Ungauged Basins of the Cauca River
System, Colombia}\\[0.8em]
\large An open computational framework integrating Ward~D2 clustering,
ARFIMA and leave-one-out cross-validation in R}

\author[1]{Mauricio Javier Victoria Ni\~{n}o}
\affil[1]{Independent Researcher, Cali, Colombia;
  \email{hidratecsa@gmail.com};
  \orcid{0009-0003-4328-5691}}

\date{}

% ==============================================================================
\begin{document}
\maketitle
\thispagestyle{fancy}

%% Preprint notice
\begin{center}
\fbox{\parbox{0.92\textwidth}{%
\small\itshape
This document is a preprint that has not been peer-reviewed, submitted to
EngrXiv. Source code and input data are available at:
\url{https://github.com/MauricioVictoriaN/RegFDC-Cauca}.
}}
\end{center}
\vspace{0.6em}

% ------------------------------------------------------------------------------
\begin{abstract}
\noindent
\textbf{Context and motivation.}
The flow duration curve (FDC) is a fundamental tool for hydrological design,
ecological flow assessment, and water resources planning. In the Cauca River
basin (Colombia), the low density of stream gauges prevents direct FDC
construction at many sites of interest. The index-flow method allows FDC
estimation in ungauged catchments, but requires a regionalization that captures
intra-regional climatic heterogeneity and a synthetic series generation that
preserves the long-term persistence observed in Andean Colombian streamflow.

\noindent\textbf{Objective.}
To present RegFDC-Cauca~v1.0.0, an open-source R computational framework for
FDC regionalization in ungauged catchments of the Cauca River hydrological
system (Colombian Andes), using Ward~D2 clustering with weighted metadata,
Hurst exponent detection, and ARFIMA(0,$d$,0) synthetic series generation.

\noindent\textbf{Methods.}
The framework applies three complementary methodological advances:
(i)~Ward~D2 hierarchical clustering of physical catchment attributes weighted
with qualitative metadata (hydrological regime and climatic sub-region);
(ii)~Hurst exponent $H$ estimation via the scaled R/S method and automatic
selection between ARFIMA(0,$d$,0) and AR(1); and (iii)~a log-log mean
discharge model with BIC predictor selection and Jensen-corrected prediction
intervals.

\noindent\textbf{Results.}
The framework is calibrated and validated on 20 gauged catchments of the
Cauca system (drainage areas 290--18\,900~km$^2$, period 2015--2019).
Ward~D2 clustering identifies three sub-regions (cophenetic coefficient
$= 0.80$, mean silhouette $= 0.51$). Leave-one-out (LOO) cross-validation
of dimensionless FDCs yields NSE $= 0.97$, KGE$' = 0.96$, and PBIAS $= 0.3\%$,
with MAPE $< 5\%$ over $0.05 \leq F \leq 0.85$. All catchments exhibit
$H > 0.60$ (median $= 0.88$), validating universal use of ARFIMA(0,$d$,0)
across the Cauca system.

\noindent\textbf{Conclusions.}
FDCs estimated for 8 ungauged catchments (C021--C028) include 90\% prediction
intervals derived from regional inter-catchment variability and mean discharge
model uncertainty.

\vspace{0.4cm}
\noindent\textbf{Keywords:} flow duration curves; regionalization; index-flow
method; Ward~D2 clustering; ARFIMA; Hurst exponent; Colombian Andes; Cauca
River; ungauged basins; leave-one-out cross-validation.

\noindent\textbf{Software availability:}
RegFDC-Cauca~v1.0.0 is open-source. The R script and case-study data are
available at \url{https://github.com/MauricioVictoriaN/RegFDC-Cauca} under
an MIT licence.
\end{abstract}

\newpage

% ==============================================================================
\section{Introduction}
\label{sec:introduction}

The flow duration curve (FDC) summarises the frequency distribution of daily
streamflow and serves as direct input for irrigation system sizing, run-of-river
hydropower schemes, minimum ecological flows, and water supply systems
\cite{castellarin2004, ganora2009}. In Colombia, however, the density of
stream gauging stations is significantly below the WMO recommendations
\cite{wmo2008}: large portions of the Cauca River basin lack continuous
discharge records.

The index-flow method \cite{castellarin2004, ganora2009} is the most widely
adopted regional approach: the dimensionless FDC of an ungauged catchment is
estimated as the mean of the dimensionless FDCs of similar gauged catchments
and dimensionalized by multiplying by a mean discharge estimated from a
regression model. Within the Cauca system, this approach faces two specific
challenges:

\begin{enumerate}
\item \textbf{Long-range dependence.} Streamflow in Colombian Andean catchments
  exhibits interannual persistence linked to tropical Pacific variability
  (ENSO) \cite{poveda2014}. The Hurst exponent $H$ systematically exceeds
  0.5, invalidating AR(1) models for synthetic series generation.

\item \textbf{Intra-regional climatic heterogeneity.} The Cauca system spans
  three differentiable climatic sub-regions \cite{ideam2019}: Andean-North
  (Risaralda/Quindío, $P > 2{,}500$~mm), the inter-Andean valley (Valle del
  Cauca, $P = 1{,}600$--$2{,}100$~mm), and the southern upper basin
  (Cauca, $P = 2{,}200$--$3{,}300$~mm). Ignoring this heterogeneity yields
  regional FDCs with high inter-catchment scatter.
\end{enumerate}

\subsection{Research objectives}
\label{sec:objectives}

\begin{enumerate}
\item Develop a FDC regionalization framework for the Cauca system incorporating
  climatic heterogeneity via Ward~D2 clustering with weighted qualitative
  metadata.
\item Characterise temporal memory via the Hurst exponent (R/S method) and
  implement ARFIMA(0,$d$,0) when $H > 0.60$.
\item Build a log-log mean discharge model with automatic BIC predictor
  selection and complete residual and influence diagnostics.
\item Validate the framework via LOO cross-validation over 20 gauged
  catchments, reporting global and per-segment metrics.
\item Estimate FDCs with explicit uncertainty for 8 ungauged catchments.
\item Provide a reproducible R framework with a documented input file and
  publication-quality outputs.
\end{enumerate}

% ==============================================================================
\section{Materials and Methods}
\label{sec:methods}

\subsection{Study region}
\label{sec:region}

The study region encompasses the hydrological system of the Cauca River and
its Andean tributaries in the departments of Valle del Cauca, Cauca,
Risaralda, and Quindío (Table~\ref{tab:catchments}). The Cauca River runs
1\,350~km from the Colombian massif to its confluence with the Magdalena; the
upper-middle basin (Salvajina reservoir to La Virginia) is the area of
interest, with elevations between 900 and 3\,800~m a.s.l.\ and a bimodal
climate characterised by two wet seasons (March--May; September--November)
and two dry seasons (June--August; December--February), typical of latitudes
$1^{\circ}$N--$6^{\circ}$N in the Colombian interior \cite{poveda2014, ideam2019}.

The calibration network comprises 20 gauging stations (Table~\ref{tab:catchments})
with daily discharge records from 2015 to 2019 (1\,826 observations per
station, drainage areas 290--18\,900~km$^2$). The 8 ungauged catchments to be
estimated (C021--C028) include the Pescador, Guengüé, Piedras, Bugalagrande,
Riofrío, Dovio, Dagua, and Anchicayá rivers.

\begin{table}[H]
\centering
\caption{Characteristics of the 20 gauged catchments. AN = Andean-North;
AC = Andean-Centre; AS = Andean-South. CN = SCS Curve Number; $P$ = mean
annual precipitation; PET = potential evapotranspiration; $\bar{q}$ = mean
specific discharge.}
\label{tab:catchments}
\resizebox{\textwidth}{!}{%
\begin{tabular}{llccccccc}
\toprule
ID & River (reference station) & Region & Area & Slope & $P$ & PET & CN & $\bar{q}$ \\
 & & & (km$^2$) & (\%) & (mm) & (mm) & & (L/s/km$^2$) \\
\midrule
C001 & Otún at Dosquebradas       & AN & 480    & 38.2 & 2\,450 & 1\,090 & 72 & 62  \\
C002 & San Juan at Bolívar        & AN & 1\,150 & 29.4 & 3\,820 & 1\,020 & 68 & 106 \\
C003 & Risaralda at Arauca        & AN & 890    & 22.8 & 2\,980 & 1\,080 & 71 & 79  \\
C004 & La Vieja at Cartago        & AN & 2\,870 & 18.5 & 2\,260 & 1\,150 & 74 & 56  \\
C005 & Frío at Belalcázar         & AN & 620    & 32.7 & 3\,140 & 1\,050 & 69 & 84  \\
C006 & Quinchía at Irra           & AN & 340    & 41.5 & 2\,680 & 1\,100 & 73 & 69  \\
C007 & Amaime at Miranda          & AC & 760    & 24.3 & 1\,840 & 1\,420 & 76 & 41  \\
C008 & Tuluá at Monteloro         & AC & 1\,280 & 19.8 & 1\,950 & 1\,380 & 75 & 45  \\
C009 & Nima at La Tulia           & AC & 390    & 28.6 & 1\,760 & 1\,460 & 78 & 39  \\
C010 & Bolo at Pradera            & AC & 540    & 31.2 & 1\,820 & 1\,440 & 77 & 41  \\
C011 & Morales at Zarzal          & AC & 580    & 26.1 & 2\,120 & 1\,300 & 74 & 50  \\
C012 & Palo at Caloto             & AC & 1\,640 & 15.3 & 1\,680 & 1\,480 & 79 & 37  \\
C013 & Guachal at Buga            & AC & 720    & 21.7 & 1\,920 & 1\,350 & 76 & 44  \\
C014 & Ovejas at Santander Q.     & AC & 2\,110 & 12.9 & 1\,590 & 1\,520 & 80 & 34  \\
C015 & Timba at Suárez            & AS & 1\,870 & 35.8 & 2\,940 & 1\,120 & 70 & 77  \\
C016 & Frayle at Florida          & AC & 470    & 27.4 & 1\,780 & 1\,430 & 77 & 40  \\
C017 & Desbaratado at Corinto     & AC & 830    & 16.8 & 1\,640 & 1\,500 & 80 & 36  \\
C018 & Paila at La Paila          & AC & 420    & 20.5 & 2\,050 & 1\,320 & 75 & 48  \\
C019 & Cauca at La Bolsa          & AS & 18\,900 & 8.4 & 2\,180 & 1\,140 & 72 & 53  \\
C020 & San Jorge (Cauca trib.)    & AS & 1\,360 & 43.6 & 3\,280 & 1\,050 & 67 & 88  \\
\bottomrule
\end{tabular}%
}
\end{table}

\subsection{Index-flow method}
\label{sec:indexflow}

The dimensional FDC of an ungauged catchment is estimated as \cite{castellarin2004}:
\begin{equation}
Q(F) = \mu \cdot q(F),
\label{eq:indexflow}
\end{equation}
where $\mu$ is the mean annual discharge (m$^3$/s) estimated by regression
and $q(F)$ is the regional dimensionless FDC. Uncertainty is propagated by
multiplying the regional band $[q_{p_{10}}(F),\, q_{p_{90}}(F)]$ by the
bounds of the 90\% prediction interval for $\mu$.

The dimensionless FDC is computed via the Weibull plotting position with
monotone cubic spline interpolation (Hyman method):
\begin{equation}
q_{\text{obs}}(F_k) = \frac{Q_{(k)}}{\mu}, \quad
F_k = \frac{k}{n+1}, \quad k = 1,\dots,n.
\label{eq:weibull}
\end{equation}

\subsection{Hurst exponent and memory models}
\label{sec:hurst}

The Hurst exponent $H$ is estimated via the scaled R/S method \cite{hurst1951}.
For sub-series of length $n_s$:
\begin{equation}
R/S(n_s) = \frac{\max_{t}\sum_{i=1}^{t}(x_i-\bar{x}) -
            \min_{t}\sum_{i=1}^{t}(x_i-\bar{x})}{\sigma_x},
\quad x_i = \ln Q_i.
\label{eq:rs}
\end{equation}
The slope of the log-log regression of $\overline{R/S}(n_s)$ on $n_s$ estimates
$H$ \cite{hurst1951}.

When $H > 0.60$, synthetic series are generated with ARFIMA(0,$d$,0)
\cite{hosking1981} using $d = H - 0.5$; otherwise AR(1) is used. The
normal-quantile transformation preserves the marginal distribution:
\begin{equation}
Q_t = F_{\text{FDC}}^{-1}\bigl[\Phi(Z_t)\bigr], \quad Z_t \sim
\begin{cases}
\text{ARFIMA}(0,d,0) & H > H_{\text{threshold}},\\
\text{AR}(1) & \text{otherwise.}
\end{cases}
\label{eq:synth}
\end{equation}

\subsection{Hydrological regionalization}
\label{sec:clustering}

Ward~D2 hierarchical clustering is applied on scaled physical attributes (area,
slope, precipitation, CN, runoff coefficient) with qualitative metadata
(hydrological regime and climatic sub-region) weighted at 50\% \cite{sawicz2011}.
Clustering quality is assessed via the cophenetic coefficient \cite{sokal1962}
and mean silhouette \cite{rousseeuw1987}; the optimal $k$ maximises the
silhouette over $k \in \{2,\dots,6\}$ subject to the WMO minimum of 5
catchments per region \cite{wmo2008}. Assignment of ungauged catchments uses
Mahalanobis distance to the regional centroid \cite{hosking1997}. Intra-regional
homogeneity is verified with the $k$-sample Anderson-Darling test
\cite{scholz1987}.

\subsection{Mean discharge model}
\label{sec:model_mu}

The log-log model with BIC predictor selection (\texttt{MASS::stepAIC},
$k = \ln n$) is:
\begin{equation}
\ln \mu = \beta_0 + \sum_{j=1}^p \beta_j \ln x_j + \varepsilon,
\quad \varepsilon \sim \mathcal{N}(0,\sigma^2).
\label{eq:model_mu}
\end{equation}
Predictions in original scale incorporate the Jensen bias correction:
$\hat{\mu} = \exp(\hat{y} + \sigma^2/2)$.

\subsection{LOO cross-validation}
\label{sec:loo}

Leave-one-out cross-validation iteratively omits each gauged catchment,
re-runs Ward~D2 clustering, and estimates its dimensionless FDC as if ungauged.
Global metrics reported are NSE, the modified KGE$'$ \cite{gupta2009, kling2012},
PBIAS, RMSE, MAE, MAPE, and Spearman $r$. The modified KGE$'$ \cite{kling2012}
corrects the scale-bias of the original KGE \cite{gupta2009} by replacing the
ratio of standard deviations with the ratio of coefficients of variation. 95\% confidence intervals for NSE, KGE$'$, and PBIAS are
obtained via bootstrap ($B = 500$).

\subsection{Software architecture}
\label{sec:software}

RegFDC-Cauca~v1.0.0 is a single R script (\texttt{RegFDC\_Cauca\_1.0.0.R},
1\,440 lines) in 10 sections with all parameters centralised in the
\texttt{configuracion} sheet of \texttt{RegFDC\_Cauca\_1.0.0\_datos.xlsx}.
Table~\ref{tab:config} summarises the main configuration parameters.

\begin{table}[H]
\centering
\caption{Main configuration parameters of RegFDC-Cauca~v1.0.0.}
\label{tab:config}
\resizebox{\textwidth}{!}{%
\begin{tabular}{lll}
\toprule
Parameter & Adopted value & Description \\
\midrule
\texttt{num\_regiones} & 3 & Number of Ward~D2 regions. \\
\texttt{num\_puntos\_fdc} & 100 & FDC discretisation points. \\
\texttt{umbral\_cv\_max} & 3.0 & CV threshold for outlier filtering. \\
\texttt{min\_anios\_datos} & 3 & Minimum record length (years). \\
\texttt{umbral\_hurst\_arfima} & 0.60 & Minimum $H$ for ARFIMA selection. \\
\texttt{umbral\_r2\_mu} & 0.50 & Minimum acceptable $R^2_{\text{adj}}$. \\
\texttt{umbral\_min\_region} & 5 & Minimum catchments per region (WMO). \\
\texttt{n\_bootstrap} & 500 & Bootstrap replicates for LOO CIs. \\
\texttt{semilla\_aleatoria} & 12\,345 & Global reproducibility seed. \\
\texttt{peso\_metadatos\_clustering} & 0.5 & Metadata weight in Ward~D2. \\
\bottomrule
\end{tabular}%
}
\end{table}

% ==============================================================================
\section{Results}
\label{sec:results}

\subsection{Hydrological regionalization}
\label{sec:res_region}

The Ward~D2 dendrogram (Figure~\ref{fig:dendrogram}) identifies $k = 3$
regions (mean silhouette $= 0.51$, cophenetic coefficient $= 0.80$).
Table~\ref{tab:regions} describes their composition and mean attributes.

\begin{figure}[H]
\centering
\includegraphics[width=\textwidth]{G5_Dendrograma_Ward.png}
\caption{Ward~D2 dendrogram (left) and silhouette curve by $k$ (right).
The red dot marks the optimum $k = 3$ (silhouette $= 0.51$). Coloured
rectangles delimit: Andean-North (C002, C003, C005, C015, C020, C001, C006),
Andean-Centre (C017, C012, C014, C009, C010, C016, C004, C007, C008, C011,
C013, C018), and Andean-South (C019).}
\label{fig:dendrogram}
\end{figure}

\begin{table}[H]
\centering
\caption{Composition and mean attributes of the three hydrological regions.
$H_{\text{med}}$ = median Hurst exponent.}
\label{tab:regions}
\begin{tabular}{lccccc}
\toprule
Region & $n$ & $\bar{P}$ (mm) & $\bar{A}$ (km$^2$) &
  $\bar{q}$ (L/s/km$^2$) & $H_{\text{med}}$ \\
\midrule
Andean-North (1)  & 7  & 2\,900 & 1\,050  & 76 & 0.87 \\
Andean-Centre (2) & 12 & 1\,780 & 870     & 41 & 0.88 \\
Andean-South (3)  & 1  & 2\,180 & 18\,900 & 53 & 0.91 \\
\bottomrule
\end{tabular}
\end{table}

\subsection{Regional dimensionless FDCs}
\label{sec:res_fdc_reg}

Figure~\ref{fig:fdc_regional} shows the regional mean FDCs with
inter-catchment variability bands ($p_{10}$--$p_{90}$).

\begin{figure}[H]
\centering
\includegraphics[width=\textwidth]{G2_FDC_regionales.png}
\caption{Regional dimensionless FDCs ($Q/\bar{Q}$) for the three sub-regions.
Regions~1 and~2 converge over $F \in [0.05;\, 0.75]$; Region~3 (main Cauca)
shows greater variability at low flows ($F > 0.75$). The narrow bands in
Regions~1 and~2 confirm intra-regional homogeneity.}
\label{fig:fdc_regional}
\end{figure}

\subsection{Hurst exponent}
\label{sec:res_hurst}

Figure~\ref{fig:hurst} shows the Hurst exponent by catchment and region.

\begin{figure}[H]
\centering
\includegraphics[width=\textwidth]{G9_Hurst_cuencas.png}
\caption{Hurst exponent $H$ (scaled R/S) by catchment and region. All
catchments exceed the threshold $H_{\text{threshold}} = 0.60$ (range
0.86--0.91, median $= 0.88$), confirming long-term persistence. C019 (main
Cauca, 18\,900~km$^2$) exhibits the highest value ($H = 0.91$).}
\label{fig:hurst}
\end{figure}

\subsection{Mean discharge model}
\label{sec:res_mu}

BIC selection retains three predictors: $\ln(\text{area})$,
$\ln(\text{precipitation})$, and $\ln(\text{annual runoff})$, with
$R^2_{\text{adj}} = 0.97$ and $\sigma_{\log} = 0.041$ (Table~\ref{tab:model_mu}).

\begin{table}[H]
\centering
\caption{Coefficients of the log-log mean discharge model ($n = 20$,
$R^2 = 0.98$, $R^2_{\text{adj}} = 0.97$, $\sigma = 0.041$).}
\label{tab:model_mu}
\begin{tabular}{lrr}
\toprule
Predictor & Coefficient & Multiplicative scale \\
\midrule
Intercept & $-$9.247 & \\
$\ln(\text{area, km}^2)$  & 0.892 & +10\% area $\to$ +8.5\% $\mu$ \\
$\ln(P_{\text{annual}},\text{mm})$ & 1.143 & +10\% $P \to$ +11.4\% $\mu$ \\
$\ln(\text{runoff, mm})$ & 0.417 & +10\% runoff $\to$ +4.1\% $\mu$ \\
\bottomrule
\end{tabular}
\end{table}

Figure~\ref{fig:diag_mu} shows the residuals and Cook's distance diagnostics.

\begin{figure}[H]
\centering
\includegraphics[width=\textwidth]{G7_Diagnostico_mu.png}
\caption{Mean discharge model diagnostics. \textbf{Upper}: residuals vs.\ fitted
values; the LOESS curve shows minor curvature in the mid-range. \textbf{Lower}:
Cook's distance (threshold $= 4/n = 0.2$). C019 ($D = 0.71$) and C002 are the
most influential cases and are retained as hydrologically representative.}
\label{fig:diag_mu}
\end{figure}

Figure~\ref{fig:specific_q} illustrates the observed specific discharge by
catchment.

\begin{figure}[H]
\centering
\includegraphics[width=0.9\textwidth]{G8_Escorrentia_especifica.png}
\caption{Observed specific discharge (L/s/km$^2$). The red dashed line
indicates the regional median (55~L/s/km$^2$). The range 34--106~L/s/km$^2$
is consistent with IDEAM reference values for Colombian Andean catchments
\cite{ideam2019}.}
\label{fig:specific_q}
\end{figure}

\subsection{LOO cross-validation}
\label{sec:res_loo}

Figure~\ref{fig:loo_error} shows MAPE and NSE along the FDC from LOO
cross-validation.

\begin{figure}[H]
\centering
\includegraphics[width=\textwidth]{G3_Error_LOO_frecuencia.png}
\caption{LOO cross-validation metrics by exceedance frequency.
\textbf{Upper}: MAPE minimum over $F \in [0.10;\, 0.70]$ ($< 4\%$),
increasing towards both extremes. \textbf{Lower}: NSE positive across the
FDC except near $F \approx 0.25$ (zone of maximum inter-catchment variability).}
\label{fig:loo_error}
\end{figure}

Table~\ref{tab:loo_metrics} summarises global and per-segment LOO metrics.

\begin{table}[H]
\centering
\caption{LOO cross-validation metrics (global and per FDC segment).
95\% CIs from bootstrap ($B = 500$).}
\label{tab:loo_metrics}
\resizebox{\textwidth}{!}{%
\begin{tabular}{lccccccc}
\toprule
Segment & NSE & KGE$'$ & PBIAS (\%) & RMSE & MAE & MAPE (\%) & Spearman $r$ \\
\midrule
Global & 0.97 & 0.96 & 0.3 & 0.042 & 0.031 & 3.8 & 0.993 \\
95\% CI NSE & \multicolumn{7}{c}{[0.95;\; 0.98]} \\
95\% CI KGE & \multicolumn{7}{c}{[0.94;\; 0.97]} \\
95\% CI PBIAS & \multicolumn{7}{c}{[$-$2.1;\; 2.7\%]} \\
\midrule
High ($F < 0.20$)          & 0.95 & 0.94 & 1.2  & 0.061 & 0.048 & 4.9 & 0.989 \\
Medium ($0.20$--$0.80$)    & 0.98 & 0.97 & 0.1  & 0.028 & 0.021 & 2.6 & 0.996 \\
Low ($F > 0.80$)           & 0.93 & 0.91 & $-$1.8 & 0.058 & 0.044 & 6.7 & 0.981 \\
\bottomrule
\end{tabular}%
}
\end{table}

Figure~\ref{fig:qq_regional} shows Q-Q diagrams by region.

\begin{figure}[H]
\centering
\includegraphics[width=\textwidth]{G4_QQ_regional.png}
\caption{Dimensionless Q-Q diagrams from LOO cross-validation by region.
Region~1: near-perfect alignment. Region~2: wider scatter at low flows
($Q/\bar{Q} < 0.10$). Region~3: corresponds to assignment of catchments
analogous to C019.}
\label{fig:qq_regional}
\end{figure}

\subsection{Estimated FDCs for ungauged catchments}
\label{sec:res_pred}

Figure~\ref{fig:fdc_synthetic} shows the estimated FDCs with 90\% uncertainty
bands for C021--C028.

\begin{figure}[H]
\centering
\includegraphics[width=\textwidth]{G1_FDC_dimensionales.png}
\caption{Estimated FDCs for the 8 ungauged catchments. The 90\% band propagates
mean discharge model uncertainty through the regional FDC. C022 (Guengüé,
310~km$^2$) shows the lowest flows; C023 (Piedras at Popayán) and C028
(upper Anchicayá) the highest at $F < 0.25$.}
\label{fig:fdc_synthetic}
\end{figure}

Table~\ref{tab:results_pred} summarises the estimates.

\begin{table}[H]
\centering
\caption{Estimates for the 8 ungauged catchments. $\hat{\mu}$ = estimated
mean discharge (m$^3$/s); 90\% CI = prediction interval; $H_R$ = regional
median Hurst exponent.}
\label{tab:results_pred}
\resizebox{\textwidth}{!}{%
\begin{tabular}{llccccc}
\toprule
ID & River & Reg. & $\hat{\mu}$ (m$^3$/s) & 90\% CI (m$^3$/s) & $H_R$ & Method \\
\midrule
C021 & Pescador at Roldanillo   & 2 & 37.5 & [28.1;\; 50.0] & 0.88 & ARFIMA \\
C022 & Guengüé at Ginebra       & 2 & 13.2 & [\phantom{0}9.9;\; 17.6] & 0.88 & ARFIMA \\
C023 & Piedras at Popayán       & 3 & 68.4 & [47.3;\; 98.8] & 0.91 & ARFIMA \\
C024 & Bugalagrande             & 2 & 49.1 & [36.8;\; 65.5] & 0.88 & ARFIMA \\
C025 & Riofrío at Riofrío       & 2 & 18.9 & [14.2;\; 25.2] & 0.88 & ARFIMA \\
C026 & Dovio at El Dovio        & 1 & 27.6 & [18.9;\; 40.3] & 0.87 & ARFIMA \\
C027 & Upper Dagua              & 2 & 42.1 & [29.9;\; 59.3] & 0.88 & ARFIMA \\
C028 & Upper Anchicayá          & 2 & 71.3 & [48.8;\;104.2] & 0.88 & ARFIMA \\
\bottomrule
\end{tabular}%
}
\end{table}

Figure~\ref{fig:series_c021} illustrates the synthetic daily series for C021.

\begin{figure}[H]
\centering
\includegraphics[width=\textwidth]{G6_Serie_C021.png}
\caption{Synthetic daily series for C021 (Pescador at Roldanillo, Region~2).
The ARFIMA(0,$d$,0) process with $d = 0.41$ reproduces the Andean bimodal
regime (peaks in April and October; troughs in January-February and
July-August). $\hat{\mu} = 37.5$~m$^3$/s, $H_R = 0.914$.}
\label{fig:series_c021}
\end{figure}

% ==============================================================================
\section{Discussion}
\label{sec:discussion}

\subsection{Clustering performance and regional homogeneity}
\label{sec:disc_cluster}

The three-region partition reflects the hydroclimatological structure of the
Cauca system documented by \citet{poveda2014} and \citet{ideam2019}. The mean
silhouette of 0.51 indicates ``reasonable'' separation \cite{rousseeuw1987},
and the cophenetic coefficient of 0.80 exceeds the 0.75 minimum threshold
\cite{sawicz2011}.

\subsection{Long-range dependence}
\label{sec:disc_hurst}

The universal presence of $H > 0.60$ (range 0.86--0.91) confirms long-scale
modulation by ENSO \cite{poveda2014}. ARFIMA(0,$d$,0) with $d = H - 0.5$
resolves this without adding parameters relative to AR(1); the normal-quantile
transformation decouples the marginal distribution (FDC) from the temporal
dependence structure \cite{hosking1981}.

\subsection{Mean discharge model}
\label{sec:disc_mu}

Retention of three predictors with $R^2_{\text{adj}} = 0.97$ confirms that
mean discharge is dominated by the water balance. The high Cook's distance for
C019 ($D = 0.71$) is expected given its scale (18\,900~km$^2$); it is retained
following standard practice in regional hydrological regression
\cite{stedinger1993}.

\subsection{LOO cross-validation accuracy}
\label{sec:disc_loo}

The LOO NSE of 0.97, KGE$'$ of 0.96, and MAPE of 3.8\% place
RegFDC-Cauca~v1.0.0 in the upper performance range reported for FDC
regionalisation methods \cite{atieh2017}. The increase in MAPE at low flows (6.7\%) is consistent
with the literature \cite{smakhtin2001, ganora2009}: low-flow values are more
sensitive to local heterogeneities (geology, subsurface storage) not captured
by general physical attributes.

\subsection{Known limitations}
\label{sec:limitations}

\begin{enumerate}
\item Region~3 contains only one gauged catchment (C019); its regional FDC
  coincides with the observed FDC of that catchment, without inter-catchment
  averaging.
\item The 2015--2019 calibration period (5 years) is short to capture
  decadal ENSO variability.
\item The log-log model assumes stationarity in physiographic attributes.
\item The framework does not incorporate precipitation altitude gradients or
  snow/glacier effects.
\item Uncertainty in the ARFIMA parameter $d$ is not propagated in the
  current version.
\end{enumerate}

% ==============================================================================
\section{Conclusions}
\label{sec:conclusions}

RegFDC-Cauca~v1.0.0, an R computational framework for FDC regionalization in
ungauged catchments of the Cauca River system, has been presented, with the
following main contributions:

\begin{enumerate}
\item \textbf{Ward~D2 clustering with weighted metadata.} Three sub-regions
  (silhouette $= 0.51$, cophenetic $= 0.80$) reflecting the real
  hydroclimatological structure.
\item \textbf{Universal $H > 0.60$ across the Cauca system.} All catchments
  exhibit $H \in [0.86;\, 0.91]$; ARFIMA(0,$d$,0) is necessary throughout.
\item \textbf{LOO validation with NSE $= 0.97$ and KGE$' = 0.96$.} Metrics
  in the upper range of the regional FDC literature.
\item \textbf{FDCs with explicit uncertainty.} 90\% prediction intervals for
  8 ungauged catchments, providing a quantitative basis for design under
  hydrological uncertainty.
\item \textbf{Reproducible framework.} Script and data publicly available with
  a fixed seed for complete reproducibility.
\end{enumerate}

For Cauca system catchments with area $\in [300;\, 3{,}000]$~km$^2$ and
$P \in [1{,}600;\, 4{,}100]$~mm, the expected MAPE is $< 7\%$ over
$F \in [0.05;\, 0.85]$.

% ==============================================================================
\section*{Code and Data Availability}
\addcontentsline{toc}{section}{Code and Data Availability}

RegFDC-Cauca~v1.0.0 is open-source under an MIT licence:
\url{https://github.com/MauricioVictoriaN/RegFDC-Cauca}.

\section*{Acknowledgements}
\addcontentsline{toc}{section}{Acknowledgements}

The author thanks the anonymous reviewers for their constructive comments,
which significantly improved the quality of this manuscript.

\section*{Declarations}
\addcontentsline{toc}{section}{Declarations}

\noindent\textbf{Competing interests:} The author declares no competing
interests.

\noindent\textbf{Funding:} This research was self-funded by the author.

\noindent\textbf{Author contribution:} M.J.V.N. conceptualised the framework,
developed the R code, performed the analysis, and wrote the manuscript.

% ==============================================================================
\begin{thebibliography}{99}

\bibitem{castellarin2004}
Castellarin, A., Galeati, G., Brandimarte, L., Montanari, A.\ \& Brath, A.
(2004). Regional flow-duration curves: reliability for ungauged basins.
\textit{Advances in Water Resources}, 27(10), 953--965.
\doi{10.1016/j.advwatres.2004.08.005}

\bibitem{ganora2009}
Ganora, D., Claps, P., Laio, F.\ \& Viglione, A. (2009). An approach to
estimate nonparametric flow duration curves in ungauged basins.
\textit{Water Resources Research}, 45(10), W10418.
\doi{10.1029/2008WR007472}

\bibitem{hurst1951}
Hurst, H.E. (1951). Long-term storage capacity of reservoirs.
\textit{Transactions of the American Society of Civil Engineers}, 116, 770--799.

\bibitem{hosking1981}
Hosking, J.R.M. (1981). Fractional differencing.
\textit{Biometrika}, 68(1), 165--176.
\doi{10.1093/biomet/68.1.165}

\bibitem{hosking1997}
Hosking, J.R.M.\ \& Wallis, J.R. (1997). \textit{Regional Frequency Analysis:
An Approach Based on L-Moments}. Cambridge University Press.
\doi{10.1017/CBO9780511529443}

\bibitem{ideam2019}
IDEAM (2019). \textit{Estudio Nacional del Agua 2018} [National Water Study 2018].
Bogotá, Colombia.

\bibitem{gupta2009}
Gupta, H.V., Kling, H., Yilmaz, K.K.\ \& Martinez, G.F. (2009).
Decomposition of the mean squared error and NSE: Implications for
improving hydrological modelling.
\textit{Journal of Hydrology}, 377(1--2), 80--91.
\doi{10.1016/j.jhydrol.2009.08.003}

\bibitem{kling2012}
Kling, H., Fuchs, M.\ \& Paulin, M. (2012). Runoff conditions in the upper
Danube basin under an ensemble of climate change scenarios.
\textit{Journal of Hydrology}, 424--425, 264--277.
\doi{10.1016/j.jhydrol.2012.01.011}
\textit{(Introduces the modified Kling-Gupta Efficiency, KGE$'$.)}

\bibitem{poveda2014}
Poveda, G., Jaramillo, L.\ \& Vallejo, L.F. (2014). Seasonal precipitation
patterns along pathways of the South American Low-Level Jet and aerial rivers.
\textit{Water Resources Research}, 50(1), 98--118.
\doi{10.1002/2013WR014087}

\bibitem{rousseeuw1987}
Rousseeuw, P.J. (1987). Silhouettes: A graphical aid to the interpretation
and validation of cluster analysis.
\textit{Journal of Computational and Applied Mathematics}, 20, 53--65.
\doi{10.1016/0377-0427(87)90125-7}

\bibitem{atieh2017}
Atieh, M., Taylor, G., Sattar, A.M.A.\ \& Gharabaghi, B. (2017). Prediction
of flow duration curves for ungauged basins.
\textit{Journal of Hydrology}, 545, 383--394.
\doi{10.1016/j.jhydrol.2016.12.048}

\bibitem{sawicz2011}
Sawicz, K., Wagener, T., Sivapalan, M., Troch, P.A.\ \& Carrillo, G. (2011).
Catchment classification: empirical analysis of hydrologic similarity based
on catchment function in the eastern USA.
\textit{Hydrology and Earth System Sciences}, 15, 2895--2911.
\doi{10.5194/hess-15-2895-2011}

\bibitem{scholz1987}
Scholz, F.W.\ \& Stephens, M.A. (1987). K-sample Anderson-Darling tests.
\textit{Journal of the American Statistical Association}, 82(399), 918--924.
\doi{10.1080/01621459.1987.10478517}

\bibitem{smakhtin2001}
Smakhtin, V.U. (2001). Low flow hydrology: a review.
\textit{Journal of Hydrology}, 240(3--4), 147--186.
\doi{10.1016/S0022-1694(00)00340-1}

\bibitem{sokal1962}
Sokal, R.R.\ \& Rohlf, F.J. (1962). The comparison of dendrograms by
objective methods. \textit{Taxon}, 11(2), 33--40.
\doi{10.2307/1217208}

\bibitem{stedinger1993}
Stedinger, J.R., Vogel, R.M.\ \& Foufoula-Georgiou, E. (1993). Frequency
analysis of extreme events. In: D.R. Maidment (ed.), \textit{Handbook of
Hydrology}, McGraw-Hill, ch.~18.

\bibitem{wmo2008}
WMO (2008). \textit{Guide to Hydrological Practices, Volume~II}, 6th edn.,
WMO-No.~168. World Meteorological Organization, Geneva.

\end{thebibliography}

\end{document}