How a Few Banks Move the Whole Credit System
This chapter explains the transmission mechanism at the center of Mauritania’s banking evidence: how a movement that begins inside one bank can become a movement in aggregate credit. The important distinction is between a bank’s raw change in lending and the bank-specific part of that change after common conditions are removed. Understanding that distinction lets the reader see why concentration matters without assuming that every large-bank movement is automatically causal.
A shock is not simply a change
Suppose a bank increases its loan book in a particular month. That observation alone tells us very little about the source of the increase. The bank may be responding to a lower policy rate, rising money supply, inflation, exchange-rate conditions, or a change in demand shared by many lenders. It may also be responding to its own decision to relax lending standards, approve a large project, change its portfolio, or reassess the risk of an important client. These influences can produce the same visible outcome—a rise in lending—but they have different implications for how the movement should be interpreted.
The paper separates them in three layers. Common conditions are forces that affect many banks at once, such as the policy rate, inflation, M3 growth, or exchange-rate movements. Observable bank characteristics describe differences that can also shape lending, including liquidity, borrower concentration, depositor concentration, profitability, and the lending–deposit spread. Idiosyncratic movement is what remains after the study accounts for the common conditions, the selected bank characteristics, and each bank’s time-invariant fixed effect.
This last term is the paper’s working definition of a bank-level shock. It does not mean that researchers have watched a credit committee make a decision and proved that the decision caused the residual. It means that the observed movement was not explained by the factors included in the first-stage model. The residual is therefore useful as a measurement object, but it carries the assumptions and omissions of the model that produced it.
That distinction matters for Mauritania because the country’s credit is concentrated among a few dominant banks. If all banks respond in roughly the same way to a change in the policy rate, the movement is common and should be represented in the macro controls. If one large bank changes its lending after a borrower-specific reassessment, that change can be important precisely because it is not shared by every other bank. The same event inside a small institution may remain mostly local.
The paper gives several examples of what can sit inside an idiosyncratic lending movement: a change in internal risk limits, a shift in portfolio strategy, a change in the composition of the client base, or the approval of a lumpy project. It also points to large-depositor behavior, borrower-specific stress, and prudential tightening as events that can alter a bank’s willingness or capacity to extend credit. These are possible sources of the residual, not separately identified causes in the data. The distinction keeps the mechanism concrete without pretending that a regression residual comes with a label attached.
Consider a stylized example. A dominant Mauritanian bank loses funding from one of its largest depositors. The withdrawal is the trigger, but the aggregate effect depends on what the bank does next. It may draw on liquid assets and keep approving loans; it may tighten internal risk limits; or it may delay new commitments while replacing the funding. The final change in lending is the outcome. Liquidity, depositor concentration, and the bank’s existing risk position are mediators that help determine how the trigger reaches that outcome. A first-stage residual may capture the bank-specific movement after measured controls, but it does not tell us which step in this chain was decisive. Keeping trigger, mediator, and outcome separate is essential when the data do not observe every internal decision.
The same reasoning applies to a large project. Approval can create a sharp increase in a bank’s loan volume that is not synchronized with the policy rate or M3 growth. It may be a genuine bank-specific decision, but it may also reflect a client event that other banks do not share. If the bank is dominant, the loan can move system credit. If another bank finances the project instead, the aggregate result may be similar while the bank-level residual is different. Granularity concerns which institutions carry the movement, not just whether the economy eventually receives financing.
A similar logic applies to profitability, but the economic object is different. An unexpected change in net income may reflect fee income, operating costs, valuation effects, or a concentrated loan exposure. It may strengthen liquidity or capital without changing the bank’s current lending pipeline. A lending shock is closer to a direct movement in the flow of credit; a profitability shock is a movement in the earnings generated by the bank’s activities. The paper keeps the two residual series separate because treating them as interchangeable would erase the very wedge it is trying to understand.
The first-stage controls also include bank fixed effects. These effects absorb persistent differences across institutions—features that do not change month to month but could otherwise be mistaken for shocks. A bank with a permanently different business model, management structure, or customer base should not be treated as experiencing a new idiosyncratic event merely because its average lending level differs from another bank’s. Fixed effects do not solve every omitted-variable problem, but they narrow the comparison to within-bank movements over time.
The mechanism can be stated as a sequence. A bank experiences an observed change in lending. The model removes the part associated with common macroeconomic conditions, measured bank characteristics, and persistent bank differences. The remaining residual is treated as a bank-specific innovation. The question is then whether that innovation matters at the system level. The answer cannot be inferred from the first step alone. It depends on the size of the bank that produced the residual and on whether other institutions can offset it.
Why size weights keep a local movement from washing out
In a system with many similarly sized banks, an unusual lending decision at one institution may have little aggregate effect. Positive and negative idiosyncratic movements can offset one another. This is the intuition behind the law of large numbers: when many small, weakly related disturbances are averaged together, the individual disturbances tend to wash out.
Mauritania’s banking structure creates a different condition. The distribution of bank size is right-skewed, with a few institutions much larger than the rest. In that setting, the individual movements of the largest banks have more aggregate weight. The paper therefore constructs a granular shock by multiplying each bank’s residual by its share of system size in the previous period and then adding those weighted residuals across banks. The baseline size measure is total assets; the paper reports that using total loans produces robust results.
The use of the previous period’s share is important. It weights a bank by its position before the current month’s shock is observed, rather than allowing the current outcome to determine its own weight. In plain language, a bank that was already large carries more influence when it experiences an unexpected movement. The weight is not a judgment that the bank is better managed or more productive. It is a measure of how much of the system’s balance sheet is exposed to that bank’s movement.
Imagine two banks that each reduce lending by the same proportional amount after a bank-specific event. If one is a small institution and the other holds a much larger share of system assets, the unweighted average treats their residuals as equally important. The size-weighted construction does not. It asks how much of the aggregate balance sheet is attached to each movement. That is the route by which a bank-level event can become a macro-financial event without every bank changing its behavior.
This does not mean that the largest bank always determines the direction of credit. Several large banks can move in opposite directions. A small bank can also have an unusually large residual relative to its own starting point. Size weighting changes the contribution each event makes to the system measure; it does not guarantee that the weighted sum will be positive, negative, or large in every month. The mechanism is about exposure and aggregation, not about giving dominant banks an inevitable influence over every outcome.
The relationship between lending and aggregate credit also has a mechanical component. Aggregate credit is itself built from the lending positions of banks, so a size-weighted sum of bank-level lending movements will naturally be related to system-wide lending. The paper’s contribution is not to discover that adding bank lending can correlate with aggregate bank lending. It is to remove selected common and observable components, construct residual movements, compare alternative weighting rules, and examine whether the residuals at dominant banks account for the high-frequency variation left unexplained by macro variables alone.
That is why the equal-weight comparison matters. The paper constructs a placebo series by averaging the first-stage lending residuals across banks and giving each bank the same weight. This is close to a representative-bank view: an unexpected movement at a very small bank counts as much as an unexpected movement at a dominant one. If the aggregate result came merely from noisy residuals, the equal-weight series might perform similarly. If the result depends on the interaction between residual movement and the skewed size distribution, the two series should behave differently.
The comparison does not establish that size is the only feature that matters. Equal weighting changes one important part of the construction while keeping the residual framework and macro controls, making it a useful internal test rather than a complete alternative theory. Nor does a strong size-weighted relationship prove that every residual is a clean supply shock. It shows that the weighting rule is relevant to how the paper connects bank-level movement to aggregate credit.
The weight should also not be read as a causal multiplier. If a bank has a large lagged share, the construction gives its residual more influence because more of the system is exposed to that bank’s balance sheet. It does not say that a one-unit change in the bank’s residual mechanically creates a fixed number of units of new credit elsewhere. The aggregate regression is a second step that estimates how the constructed shock moves with aggregate credit after macro controls are included. That distinction becomes especially important when the aggregate outcome and the bank-level lending data are built from related balance sheets.
The arrows in this diagram should be read carefully. The macro conditions and bank characteristics are explanatory components of the observed movement; the bank-specific event is a possible source of the residual, not a directly observed label. The residual is weighted and aggregated into a predictor of credit growth. The equal-weight path is a comparison designed to test whether the size distribution carries explanatory content. None of the arrows, by itself, is a claim that the paper has identified a fully causal structural pathway.
The mechanism also has identifiable failure paths. If an omitted macroeconomic disturbance affects several banks at once, it may remain in the residual and look idiosyncratic. If the residuals are strongly correlated across banks after conditioning, the assumption that they are weakly related is weakened. If bank-size weights are measured poorly, the aggregate shock may misrepresent exposure. If the relevant response takes place over several months rather than in the same month, a contemporaneous regression may miss it. And if the largest banks are large because they face systematically different borrowers or funding conditions, size may be standing in for other differences that the controls do not fully capture.
The choice not to include time fixed effects in the first-stage panel illustrates the trade-off. Time effects could absorb more common variation, but they could also remove the monthly variation the researchers need in order to construct residual shocks. The paper instead uses explicit macro controls and bank characteristics, then checks the residual behavior with diagnostics. That is a design choice, not a guarantee that all common movements have been removed. It preserves the signal the granular exercise needs at the cost of leaving the reader responsible for considering whether the controls are sufficient.
The aggregate step is similarly limited. The paper treats the size-weighted granular series as a direct predictor of aggregate financial intermediation rather than as a conventional 2SLS instrument. A strong relationship can therefore be informative about concentration and short-run co-movement without establishing a fully causal supply shock. A later chapter will examine the fit and diagnostics in detail; the mechanism here is the map that tells us what those statistics are meant to represent.
These qualifications do not make the granular idea useless. They define the judgment it can support. The approach is well suited to asking whether a concentrated banking system’s short-run credit movements are associated with residual lending changes at its dominant institutions. It is less suited, on its own, to saying exactly which internal decision caused a residual, whether the relationship persists over a longer horizon, or whether the same coefficient would appear in another country.
For Mauritania, the central intuition is therefore precise. An average-bank model can understate the importance of a dominant lender because it treats institutions as interchangeable. A granular model preserves the difference in exposure created by the size distribution. The next task is to examine how the researchers operationalize that intuition—what they measure, what the first-stage fit means, and how much confidence the diagnostics can reasonably provide.