Anladım — kendi model dokümanınıza ekleyeceğiniz, low granular metodolojiyi kronolojik akışla anlatan tek bir bölüm istiyorsunuz. Başlık sayfası, TOC, appendix gibi standalone doküman öğeleri olmadan.
Bunu doğrudan chat'te akıcı prose olarak yazayım — kopyalayıp kendi dokümanınıza yapıştırabileceğiniz şekilde. (İsterseniz sonrasında formülleri düzgün render eden bir .docx snippet'i de hazırlarım, ama önce metnin akışını görüp onaylamanız daha mantıklı.)
Low Granular Sight Deposits — Liquidity Modelling Methodology
Purpose and context
This section describes the methodology applied to sight (non-maturity) deposits whose portfolio exhibits low granularity. A portfolio is low granular when a small number of clients hold a large share of the total outstanding, and when the unitary outstanding volatility differs strongly from one client to another. The low-granular approach is applied when 80% of the total balances are held by 20% or less of the clients, and the portfolio contains fewer than 10,000 clients — a condition typically met by non-retail deposits (Corporate Banking, Securities Services, Wealth Management).
As in the standard framework, the schedule is built on the existing client base only (new production is excluded), and two schedules are produced: Schedule A, the average run-off, and Schedule B, the run-off incorporating uncertainties.
The methodology proceeds in the following steps.
Step 1 — Isolate the Very Material Depositors (VMD)
The first operation is to remove the most material clients from the calibration scope. VMD are defined as all clients whose outstanding exceeds the materiality threshold (EUR 2 bn or 5% of total balances). They are removed because their large, highly volatile outstanding would otherwise dominate the statistics and distort the calibration.
Each VMD is studied individually: the part above the threshold is scheduled by educated guess, with a maximum maturity equal to the time needed to close the relationship (generally 6 months), while the part below the threshold is scheduled following the most volatile cluster. Once the model is live, VMD outstanding is monitored so that the educated guess can be adjusted and newly emerging VMD detected.
Step 2 — Build the volatility clusters
The remaining clients (excluding VMD) are grouped into clusters built from the outstanding standard deviation, so that clients with similar unitary volatility fall in the same cluster.
The primary method is k-means, with the number of clusters k chosen via the elbow method (in practice between 2 and 5). If the algorithm produces more than one cluster with fewer than 20 clients, k must be reduced, since closure rates cannot be reliably calibrated on very small clusters; a single small cluster that remains takes the pooled closure rates of the other clusters. Where k-means cannot be applied, the Pareto classes (three classes built from the cumulative distribution of the outstanding standard deviation) are used as a fallback, with no class holding fewer than 20 counterparts.
Step 3 — Determine the stable outstanding of each cluster
For each cluster, the outstanding is split into a stable and an unstable part. Stable balances are those remaining after accounting, at a high confidence level, for both the inter-month and intra-month volatilities of the average amount per customer; they may refinance long-term assets. The unstable balances are the remainder. For Schedule A the total balances are assumed stable; for Schedule B the two volatilities are removed as follows.
Inter-month volatility — Core Volume Ratio (CVR). The core volume is the share not sensitive to month-to-month variation, obtained by multiplying the total volume by the CVR of class i:
$$CVR = 1 - \frac{P_{85%}\big(X_{t \in [1,T]}\big)}{Average\ Balance_{,T}}$$
where $X_t$ is the absolute value of the monthly variation of the average balance per customer between t and t−1, and $P_{85%}$ its 85th percentile. Using the absolute variation makes the formula insensitive to any global trend; the 85th percentile gives a conservative estimate.
Intra-month volatility — Intra-Month stable Ratio (IMR). The IMR isolates the part that is volatile within the month. For each cluster it is the 15th percentile of the ratio of the daily outstanding to the average outstanding of the corresponding month:
$$IMR = P_{15%}\left(\frac{Daily\ Outstanding_{,t}}{Monthly\ Average\ Outstanding_{,m}}\right)$$
For a given cluster, the stable proportion of the outstanding is the product $CVR_i \times IMR_i$.
Step 4 — Amortize the stable outstanding (Monte Carlo)
The amortization of the stable outstanding is obtained by simulating account closures. For each cluster, the historical monthly closure rates are first computed (an account is considered closed at date t if its balance was non-zero before t and stays zero from t to the most recent date, with a minimum of three consecutive zero months). If monthly closures are too few or too volatile, quarterly/half-yearly rates or a single aggregated distribution are used instead.
The amortizing schedule is then computed with a Monte Carlo approach (usually N = 500 simulations). In each simulation a starting month is drawn; then, for each future month, a closure rate is drawn from the historical rates of that month and cluster, and for each client a Bernoulli variable is drawn with that rate — outcome 1 keeps the client's last outstanding, outcome 0 closes the account. Summing across clients gives one path; over N simulations:
- Schedule A is the average path;
- Schedule B is the 1st percentile of the distribution, to which an environmental uncertainty of 5% per year is then applied.
The schedule does not incorporate life-cycle effects, as the targeted client segments are not subject to them.
Step 5 — Implement in the liquidity gap
The stable/unstable shares are fixed at the level of the last available date t. The modelled stable-part share aggregated across clusters is:
$$MSP = \frac{\sum_{i=1}^{k} LO_i \times CVR_i \times IMR_i}{\sum_{i=1}^{k} LO_i}$$
where $LO_i$ is the outstanding of cluster i at the last available date. To avoid treating new, potentially non-structural cash as immediately stable, the stable outstanding is reduced by the reference ratio:
$$RR(t) = \min!\left(100%,;\ \frac{\frac{1}{12}\sum_{i=0}^{11} TO_{clusters}(t-i)}{TO_{clusters}(t)}\right)$$
When the 12-month average outstanding is below the last outstanding, the stable outstanding is reduced (new cash integrated gradually); a decline in outstanding is reflected immediately.
The liquidity schedule at date t is then built by filtering the cluster outstanding successively through RR, IMR and CVR, so that the four slices sum to the full cluster outstanding:
| Component | Amount | Liquidity schedule |
|---|---|---|
| VMD | $TO_0(t)$ | Account-by-account (educated guesses + Schedule B) |
| Unstable | $(1 - RR)\sum_j TO_j(t)$ | 1 month |
| Unstable | $(1 - IMR_i),RR\sum_j TO_j(t)$ | 1 day |
| Stable (inter-month sensitive) | $IMR_i(1 - CVR_i),RR\sum_j TO_j(t)$ | 1 month |
| Stable (core) | $IMR_i,CVR_i,RR\sum_j TO_j(t)$ | Schedule B |
Step 6 — Split the stable part into operational / non-operational
Finally, the stable part of each cluster (and of each VMD) is split, as defined for the LCR, into an operational part x and a non-operational part y ($s = x + y$), scheduled against the materiality threshold. The operational part is placed against the threshold first (the stable part is always kept longer than the unstable part):
- If $x \geq$ threshold: x fills the space below the threshold and all non-operational balances sit above it ($y_b = 0$, $y = y_a$).
- If $x <$ threshold: the non-operational part completes x up to the threshold ($y_b = \text{threshold} - x$), and the remainder is $y_a$.
The part below the threshold ($x_b$) follows the modelled Schedule B; the classic non-operational part ($y_b$) follows an educated guess; and the parts above the threshold ($x_a$, $y_a$) are scheduled linearly over the end-of-relationship tenor (generally 6 months). Entities using a structural / non-structural split apply the same logic.
Bu akış, dokümanı kronolojik olarak (VMD → clustering → stable belirleme → amortization → gap implementasyonu → operational split) sıralıyor.
İşinize yarayacak iki şeyi sorayım: (1) Numaralandırma (Step 1–6 mı, yoksa kendi dokümanınızın numaralandırmasına uyacak "X.1, X.2…" gibi bir yapı mı istersiniz?), ve (2) bunu formüller düzgün render olacak şekilde .docx parçası olarak da hazırlayayım mı, yoksa bu metin kopyalamanız için yeterli mi?
Yorumlar
Yorum Gönder