## Derivative Margin Calls: A New Driver of MMF Flows

## Source details

**Canonical URL:** [Derivative Margin Calls: A New Driver of MMF Flows](https://www.imf.org/-/media/files/publications/wp/2023/english/wpiea2023061-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2023/english/wpiea2023061-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2023/english/wpiea2023061-print-pdf.pdf.json)

---

### EMIR data scope and preprocessing
- EMIR data are transaction-by-transaction data on derivatives collected through trade repositories and reported by entities resident in the EU since February 2014.
- For each derivative transaction more than 120 data fields are available, including: type of derivative, underlying security, price, amount outstanding, execution and clearing venues, valuation, collateral (margin) and life-cycle events.
- Analysis uses the EMIR sub-sample accessible to the ECB restricted to trades reported by counterparties located in the euro area.
- Reported data are enriched with sector classification using the algorithm of Lenoci and Letizia (2021), combining ECB, EIOPA and ESMA lists and ECB’s RIAD, Orbis, Refinitiv Lipper and Bank Focus.
- Data cleaning and manipulation steps:
  - Pair the two legs of a trade where possible.
  - Run quality checks and remove outliers.
  - Remove portfolios for which either VM stock posted or VM stock received exceeds 80% of the total notional value of the portfolio (80% threshold drops less than 0.1% of notional value compared with 100%).
  - Construct a unique collateral portfolio code by concatenating LEIs of the two counterparties with reported collateral portfolio codes; if missing replace by reported VM value.
  - Assign a currency to each portfolio by converting values to EUR and defining that a portfolio receives (posts) margin in a given currency if the outstanding notional of contracts receiving (posting) margin in that currency exceeds an 80% threshold of the portfolio’s total outstanding notional.
- Data limitations: missing values, residual unpaired transactions, possible under-reporting.

### Definitions and computation of VM flows from reported stocks
- Reported VM variables are provided at portfolio level as stocks: one value for VM received and one value for VM posted for each collateral portfolio (one typically equals 0).
- Net VM stock posted at time t for a portfolio: V_t = V^P_t − V^R_t.
- Flow of margin payments on day t: flow_t = V_t − V_{t−1}.
- Separation into positive and negative parts:
  - VM flow posted: positive part of flow_t (equal to flow_t if flow_t > 0, and 0 otherwise) — signals net liquidity outflow (margin call).
  - VM flow received: absolute value of negative part of flow_t (equal to |flow_t| if flow_t < 0, and 0 otherwise) — signals net liquidity inflow.
- Positive and negative parts are non-negative by construction.
- Aggregation: positive and negative parts of flow_t are aggregated separately from individual counterparty to country-sector level to match SHSS data; aggregation before separation would cause offsetting and larger information loss.

### VM currency and EUR share
- Around 60% of the VM flows over the three-months period are in EUR.

### VM flows during March 2020 turmoil (aggregated euro area non-bank entities)
- Daily VM payment flows increased more than fivefold during the March 2020 market turmoil:
  - From around €5 bn in the first half of February 2020
  - To more than €20 bn in March 2020
  - With peaks of over €30 bn
- CCPs and banks are excluded from the non-bank VM flows analysis because CCP positions are fully balanced (VM received equals VM paid) and banks often pay VM on behalf of clients.

### Reducing the high data dimension in the model
- Investors classified into 190 investor groups (10 sectors × 19 countries).
- Regression sample restricted to investor groups with:
  - The largest holdings of MMFs in each domicile (ranked by share of MMFs’ TNAs held), restricted to the top ten investor groups per domicile.
  - Among those top ten, the five sectors with the largest VM flows are selected for each domicile.
- Overlap across domiciles yields nine investor sectors of interest.
- Rationale: selection combines investment into a MMF and the size of VM payments because VM payments depend on sector size, derivative usage and market-value volatility.

### Top holders and VM flow statistics (selected figures)
- Luxembourgish MMFs: 262 funds, 270 asset-shares, TNA: EUR 89 bn.
- French MMFs: 250 funds, 265 asset-shares, TNA: EUR 327 bn.
- Irish MMFs: 50 funds, 56 asset-shares, TNA: EUR 35 bn.
- Selected VM flow posted (daily, EUR mn) examples:
  - FR Bank: 5th percentile 170, Median 10,712, 95th percentile 27,361, Mean 13,351, Std. dev. 7,189
  - NL PF: 5th percentile 424, Median 1,626, 95th percentile 4,134, Mean 1,852, Std. dev. 1,206
  - LU IF: 5th percentile 49, Median 1,073, 95th percentile 2,337, Mean 1,175, Std. dev. 582
  - DE IF: 5th percentile 206, Median 722, 95th percentile 2,636, Mean 900, Std. dev. 729
  - FR IF: 5th percentile 198, Median 366, 95th percentile 1,350, Mean 468, Std. dev. 328
- Selected VM flow received (daily, EUR mn) examples:
  - FR Bank: 5th percentile 3,862, Median 10,099, 95th percentile 29,125, Mean 13,376, Std. dev. 8,611
  - NL PF: 5th percentile 366, Median 1,783, 95th percentile 3,831, Mean 1,862, Std. dev. 1,171
  - LU IF: 5th percentile 42, Median 1,150, 95th percentile 1,928, Mean 1,140, Std. dev. 485
  - DE IF: 5th percentile 122, Median 621, 95th percentile 1,957, Mean 805, Std. dev. 688
  - FR IF: 5th percentile 156, Median 364, 95th percentile 1,017, Mean 446, Std. dev. 305
- VM flows based on beginning of February to end of April 2020; table sorted by mean of VM flow posted and received.

### Main empirical findings — contemporaneous effects (February–April 2020)
- Estimation approach:
  - Panel regressions with standard errors clustered at the fund level.
  - Model specifications: Random Effects (RE), Fixed Effects (FE), and lagged dependent variable models.
- Sign and direction:
  - VM payments posted by investor sectors are associated with MMF outflows (investors withdraw funds to post VMs).
  - VM flows received by investor sectors are associated with MMF inflows (VM received increases MMF inflows).
- Robustness:
  - Coefficients are fairly stable across FE, RE and lagged dependent variable specifications in size and significance.
  - Hausman test results:
    - Irish MMF flows: RE model appropriate (Hausman p-value: 0.67 in outflows table; p-value: 0.81 in inflows table).
    - Luxembourgish and French MMF flows: FE model preferable (Hausman p-values: 0.08 and 0.00 respectively in outflows; p-values: 0.00 in inflows for Luxembourg and France).
- Key sectors with significant effects on outflows and inflows:
  - Dutch PFs, Luxembourgish IFs, German IFs are the top non-bank sectors with largest derivative exposures and significant effects.
  - These sectors typically use MMFs both as a source of liquidity to meet VM payments and as storage when VM is received.

### Magnitude of estimated effects and timing
- Range of significant coefficients across regressions: 0.002 to 0.037.
  - Interpretation: if a sector faced VM flows of the size of EUR 1 bn, an MMF held by this sector is estimated to experience flows between EUR 2 mn and EUR 37 mn.
  - Example: Irish MMFs — estimated coefficients for VM paid by Dutch PFs around 0.010 imply EUR 1 bn VM need → estimated MMF outflows ≈ EUR 10 mn.
- Leads/lags and timing:
  - Leads of VM flows posted:
    - Evidence limited; one finding: German IFs’ forward VM posted values have a significant impact on French MMF outflows.
    - Significant lead coefficients range up to 0.021, implying EUR 1 bn forward VM payment can trigger MMF flows up to EUR 21 mn.
  - Lags of VM flows received:
    - German IFs and Luxembourgish IFs deposit received VM funds into Luxembourgish and French MMFs with a one-day delay.
    - Significant lag coefficients range from 0.004 to 0.020, implying EUR 1 bn VM received can trigger MMF inflows between EUR 4 mn and EUR 20 mn with lag.
  - Operational and behavioral considerations: delayed reinvestment may result from late intraday arrivals, expectations of market reversal, or diversification across asset types.

### Leads/lags regression highlights (selected coefficients)
- Outflows (lagged dependent variable with leads of VM posted):
  - LU IF (t): significant coefficients 0.002** and 0.002** for Luxembourgish MMFs; 0.008** and 0.009*** for French MMFs.
  - DE IF (t): significant for Irish MMFs with coefficient 0.024** and 0.020* in some specifications.
  - NL PF (t): coefficients 0.010** and 0.011** for Luxembourgish MMFs; 0.009** and 0.009*** for French and Irish MMFs.
- Inflows (lagged dependent variable with lags of VM received):
  - DE IF (t-1): coefficients 0.005** and 0.004* for Luxembourgish and French MMFs; 0.020** and 0.017* for Irish MMFs.
  - NL PF (t): coefficients 0.007** and 0.007* for Luxembourgish and French MMFs; 0.007** and 0.006** in other specifications.
  - LU IF (t-1): coefficient 0.005** for Irish MMFs in one specification.

### Model explanatory power — aggregate shares explained (February–April 2020)
- MMF Outflows (share of total February–April 2020 outflows explained):
  - Random Effects: (1) All MMF outflows 48%; (2) Predict using significant coefs. 47%; (3) Only MMFs held by investors 63%; (4) Cases (2) and (3) combined 62%.
  - Fixed Effects: (1) 47%; (2) 45%; (3) 62%; (4) 60%.
  - Lagged Dep. Var.: (1) 37%; (2) 39%; (3) 49%; (4) 52%.
  - Best-performing model predicting all MMF outflows: RE model predicts 48% of aggregate outflows.
- MMF Inflows (share of total February–April 2020 inflows explained):
  - Random Effects: (1) All MMF inflows 27%; (2) Predict using significant coefs. 19%; (3) Only MMFs held by investors 37%; (4) Cases (2) and (3) combined 26%.
  - Fixed Effects: (1) 14%; (2) 7%; (3) 19%; (4) 9%.
  - Lagged Dep. Var.: (1) 37%; (2) 32%; (3) 50%; (4) 43%.
  - Overall, inflow models explain between 7% to 50% depending on specification and aggregation method; models for outflows perform better than for inflows.
- With leads/lags included (one lead / two leads; one lag / two lags):
  - MMF Outflows examples:
    - (1) All MMF outflows: 49% / 47% / 41%  and 47% / 44% / 40%
    - (3) Only MMFs held by investors: 65% / 62% / 54%  and 62% / 58% / 53%
  - MMF Inflows examples:
    - (1) All MMF inflows: 28% / 14% / 37%  and 29% / 15% / 37%
    - (3) Only MMFs held by investors: 38% / 19% / 50%  and 39% / 20% / 50%

### Concrete quantitative mechanics (selected estimates)
- German IFs are estimated to withdraw around €20 mn from each French MMF that they hold to post €1 bn of VM today (contemporaneous effect).
- German IFs are estimated to withdraw another €21 mn to post €1 bn of VM in two days (2-day forward estimate).
- Regarding inflows, German IFs are estimated to invest around €17 mn to €20 mn into each French MMF held by them one day after they received VM flow of €1 bn.
- On the same day of receiving VM flow of €1 bn, German IFs are estimated to invest around €4 mn to €5 mn into each Luxembourgish MMF held by them.

### Robustness, subperiods and caveats
- Results robust to different model specifications (FE, RE, lagged dependent variable) and to inclusion of leads and lags of VM flows.
- Subperiod analysis comparing February–April and March shows variation in coefficient significance across countries and models; effects during the most acute phase (March 2020 “dash for cash”) are likely larger.
- Using country-sector level VM flows (instead of firm-level) likely introduces measurement error and downward bias in estimates.
- Shortening sample to March 2020 tends to increase estimated coefficients for outflows but may reduce significance due to smaller sample size.

### Policy implications
- Non-bank liquidity management should account for the fact that the cash-like properties of MMF shares can fail under stress: MMFs can decline in value and can suspend redemptions in exceptional circumstances.
- MMFs should be made more resilient to significant outflows, and the structure of their investor base should be taken into account.
- Monitoring interconnectedness across markets and across borders is important, including from relatively small but volatile links introduced by daily margin exchange requirements.
- Daily exchange of high-quality collateral reduces counterparty credit risk but can increase liquidity risk and create spillovers across markets.
- Macro- and market-stabilizing policy actions that reduce market volatility (and thus VM payments) can help stabilise MMF outflows.

### Data, limitations, and avenues for further research
- Analysis combines three highly granular unique datasets to construct a daily fund-level panel spanning from February to April 2020.
- Limitation: analysis could not be conducted at firm-to-MMF level due to data unavailability; results are at country-sector-to-MMF level.
- Suggested further research:
  - Non-bank liquidity responses beyond MMFs, including bank deposits, the repo market and credit or liquidity lines.
  - The role of initial margin payments and redemptions as additional liquidity drains on non-banks during stress.
  - Broader interconnectedness between the derivatives market and other markets using granular regulatory datasets.

*Source: Working Paper No. WP/2023/061*

### 2.1    VM payments from EMIR data

### 2.1    VM payments from EMIR data

### EMIR data scope and preprocessing
- EMIR data are transaction-by-transaction data on derivatives collected through trade repositories and reported by entities resident in the EU since February 2014.
- For each derivative transaction more than 120 data fields are available, including: type of derivative, underlying security, price, amount outstanding, execution and clearing venues, valuation, collateral (margin) and life-cycle events.
- The paper works with EMIR data accessible to the ECB, focusing on a sub-sample restricted to trades reported by counterparties located in the euro area.
- The reported data are enriched with sector classification using the algorithm of Lenoci and Letizia (2021), which combines information from four official lists (ECB’s lists of monetary financial institutions and investment funds, EIOPA’s list of insurance undertakings and ESMA’s list of CCPs) and four other data sources (ECB’s RIAD, Orbis, Refinitiv Lipper and Bank Focus).
- Data cleaning and manipulation steps:
  - Pair the two legs of a trade where possible (data are initially reported by both counterparties).
  - Run quality checks and remove outliers.
  - Remove portfolios for which either VM stock posted or VM stock received exceeds 80% of the total notional value of the portfolio (total notional computed as the sum of notional values of all contracts in the portfolio). The 80% threshold is chosen as a conservative filter; using 80% instead of 100% drops less than 0.1% of notional value.
  - Construct a reliable and unique collateral portfolio code by concatenating the legal entity identifiers (LEIs) of the two counterparties with the reported collateral portfolio codes; if a reported portfolio code does not exist it is replaced by the reported VM value.
  - Assign a currency to each portfolio for VM posted/received by converting values to EUR and defining that a portfolio receives (posts) margin in a given currency if the outstanding notional of contracts receiving (posting) margin in that currency exceeds an 80% threshold of the portfolio’s total outstanding notional.

- Data limitations noted: missing values, residual unpaired transactions, possible under-reporting.

### Definitions and computation of VM flows from reported stocks
- Reported VM variables are provided at portfolio level as stocks: one value for VM received and one value for VM posted for each collateral portfolio (one of these typically equals 0).
- Define net VM stock posted at time t for a portfolio as V_t = V^P_t − V^R_t, where V^P_t is VM stock posted and V^R_t is VM stock received.
- Define the flow of margin payments on day t as flow_t = V_t − V_{t−1}.
- Separate positive and negative parts:
  - VM flow posted: positive part of flow_t (equal to flow_t if flow_t > 0, and 0 otherwise) — signals net liquidity outflow (margin call).
  - VM flow received: absolute value of negative part of flow_t (equal to |flow_t| if flow_t < 0, and 0 otherwise) — signals net liquidity inflow.
- Positive and negative parts are non-negative by construction.
- Aggregation: positive and negative parts of flow_t are aggregated separately from individual counterparty to country-sector level to match the level of SHSS data; aggregation before separation would cause offsetting and larger information loss.

### VM currency and EUR share
- Around 60% of the VM flows over the three-months period are in EUR.

### VM flows during March 2020 turmoil (aggregated euro area non-bank entities)
- Daily VM payment flows increased more than fivefold during the March 2020 market turmoil:
  - From around €5 bn in the first half of February 2020
  - To more than €20 bn in March 2020
  - With peaks of over €30 bn
- Note: CCPs and banks are excluded from the chart/analysis of non-bank VM flows because CCP positions are fully balanced (VM received equals VM paid) and banks often pay VM on behalf of clients.

*Italic source: wpiea2023061-print-pdf - 2.1    VM payments from EMIR data*

### 3.1    Reducing the high data dimension in the model

### 3.1    Reducing the high data dimension in the model

### Method for reducing dimensionality and investor selection
- Investors classified into 190 investor groups (10 sectors × 19 countries).
- Regression sample restricted to investor groups with:
  - The largest holdings of MMFs in each domicile (ranked by share of MMFs’ TNAs held), restricted to the top ten investor groups per domicile (Table 3).
  - Among those top ten, the five sectors with the largest VM flows are selected for each domicile.
- Overlap across domiciles yields nine investor sectors of interest (listed in Table 4).
- Rationale: selection combines investment into a MMF and the size of VM payments because VM payments depend not only on sector size but also on derivative usage and market-value volatility of those derivatives.

### Top holders and domicile-level MMF statistics (from Table 3 and related text)
- Sample counts and TNAs:
  - Luxembourgish MMFs: 262 funds, 270 asset-shares, TNA: EUR 89 bn.
  - French MMFs: 250 funds, 265 asset-shares, TNA: EUR 327 bn.
  - Irish MMFs: 50 funds, 56 asset-shares, TNA: EUR 35 bn.
- Investor groups highlighted across domiciles:
  - Luxembourgish IFs, Dutch PFs, German IFs stand out as large holders and large VM flow sectors.

### VM flow statistics by selected investor group (Table 4 — daily VM flow, EUR mn)
- VM flow posted (selected figures):
  - FR Bank: 5th percentile 170, Median 10,712, 95th percentile 27,361, Mean 13,351, Std. dev. 7,189
  - NL PF: 5th percentile 424, Median 1,626, 95th percentile 4,134, Mean 1,852, Std. dev. 1,206
  - LU IF: 5th percentile 49, Median 1,073, 95th percentile 2,337, Mean 1,175, Std. dev. 582
  - DE IF: 5th percentile 206, Median 722, 95th percentile 2,636, Mean 900, Std. dev. 729
  - FR IF: 5th percentile 198, Median 366, 95th percentile 1,350, Mean 468, Std. dev. 328
- VM flow received (selected figures):
  - FR Bank: 5th percentile 3,862, Median 10,099, 95th percentile 29,125, Mean 13,376, Std. dev. 8,611
  - NL PF: 5th percentile 366, Median 1,783, 95th percentile 3,831, Mean 1,862, Std. dev. 1,171
  - LU IF: 5th percentile 42, Median 1,150, 95th percentile 1,928, Mean 1,140, Std. dev. 485
  - DE IF: 5th percentile 122, Median 621, 95th percentile 1,957, Mean 805, Std. dev. 688
  - FR IF: 5th percentile 156, Median 364, 95th percentile 1,017, Mean 446, Std. dev. 305
- Notes: VM flows based on beginning of February to end of April 2020; table sorted by mean of VM flow posted and received.

### Main empirical findings — contemporaneous effects (February–April 2020)
- Estimation approach:
  - Panel regressions with standard errors clustered at the fund level.
  - Model specifications: Random Effects (RE), Fixed Effects (FE), and lagged dependent variable models.
- Sign and direction:
  - VM payments posted by investor sectors are associated with MMF outflows (investors withdraw funds to post VMs).
  - VM flows received by investor sectors are associated with MMF inflows (VM received increases MMF inflows).
- Robustness:
  - Coefficients are fairly stable across FE, RE and lagged dependent variable specifications in size and significance.
  - Hausman test results:
    - Irish MMF flows: RE model appropriate (Hausman p-value: 0.67 in outflows table; p-value: 0.81 in inflows table).
    - Luxembourgish and French MMF flows: FE model preferable (Hausman p-values: 0.08 and 0.00 respectively in outflows; p-values: 0.00 in inflows for Luxembourg and France).
- Key sectors with significant effects on outflows and inflows:
  - Dutch PFs, Luxembourgish IFs, German IFs are the top non-bank sectors with largest derivative exposures and significant effects.
  - These sectors typically use MMFs both as a source of liquidity to meet VM payments and as storage when VM is received.

### Magnitude of estimated effects and economic significance
- Range of significant coefficients across regressions: 0.002 to 0.037.
  - Interpretation: if a sector faced VM flows of the size of EUR 1 bn, an MMF held by this sector is estimated to experience flows between EUR 2 mn and EUR 37 mn.
  - Example: Irish MMFs — estimated coefficients for VM paid by Dutch PFs around 0.010 imply EUR 1 bn VM need → estimated MMF outflows ≈ EUR 10 mn.
- Leads/lags and timing:
  - Leads (forward values) of VM flows posted:
    - Evidence limited; one notable finding: German IFs’ forward VM posted values have a significant impact on French MMF outflows (see Table 8).
    - Significant lead coefficients range up to 0.021, implying EUR 1 bn forward VM payment can trigger MMF flows up to EUR 21 mn.
  - Lags of VM flows received:
    - German IFs and Luxembourgish IFs deposit received VM funds into Luxembourgish and French MMFs with a one-day delay (Table 9).
    - Significant lag coefficients range from 0.004 to 0.020, implying EUR 1 bn VM received can trigger MMF inflows between EUR 4 mn and EUR 20 mn with lag.
  - Operational and behavioral considerations: delayed reinvestment may result from late intraday arrivals, expectations of market reversal, or diversification across asset types.

### Model explanatory power — aggregate shares explained (Table 7)
- MMF Outflows (share of total February–April 2020 outflows explained by model):
  - Random Effects: (1) All MMF outflows 48%; (2) Predict using significant coefs. 47%; (3) Only MMFs held by investors 63%; (4) Cases (2) and (3) combined 62%.
  - Fixed Effects: (1) 47%; (2) 45%; (3) 62%; (4) 60%.
  - Lagged Dep. Var.: (1) 37%; (2) 39%; (3) 49%; (4) 52%.
  - Best-performing model predicting all MMF outflows: RE model predicts 48% of aggregate outflows.
- MMF Inflows (share of total February–April 2020 inflows explained by model):
  - Random Effects: (1) All MMF inflows 27%; (2) Predict using significant coefs. 19%; (3) Only MMFs held by investors 37%; (4) Cases (2) and (3) combined 26%.
  - Fixed Effects: (1) 14%; (2) 7%; (3) 19%; (4) 9%.
  - Lagged Dep. Var.: (1) 37%; (2) 32%; (3) 50%; (4) 43%.
  - Overall, inflow models explain between 7% to 50% depending on specification and aggregation method; models for outflows perform better than for inflows.
- Interpretation caveats:
  - Estimates are averages over February–April 2020; effects during the most acute phase (e.g., March 2020 “dash for cash”) are likely larger.
  - Using country-sector level VM flows (instead of firm-level) likely introduces measurement error and downward bias in estimates.
  - Shortening sample to March 2020 tends to increase estimated coefficients for outflows (but may reduce significance due to smaller sample size).

### Additional empirical specifics (leads/lags regression highlights)
- Table 8 (lagged dependent variable specification for outflows with leads of VM posted):
  - LU IF (t): significant coefficients 0.002** and 0.002** for Luxembourgish MMFs; 0.008** and 0.009*** for French MMFs (p-values reported in table).
  - DE IF (t): significant for Irish MMFs with coefficient 0.024** and 0.020* in some specifications.
  - NL PF (t): coefficients 0.010** and 0.011** for Luxembourgish MMFs; 0.009** and 0.009*** for French and Irish MMFs.
- Table 9 (lagged dependent variable specification for inflows with lags of VM received):
  - DE IF (t-1): coefficients 0.005** and 0.004* for Luxembourgish and French MMFs; 0.020** and 0.017* for Irish MMFs.
  - NL PF (t): coefficients 0.007** and 0.007* for Luxembourgish and French MMFs; 0.007** and 0.006** in other specifications.
  - LU IF (t-1): coefficient 0.005** for Irish MMFs in one specification.

*Source: wpiea2023061-print-pdf - 3.1    Reducing the high data dimension in the model*

### 0.021 and comes in addition to the significant estimate of the contemporary effect of similar

### Derivative Margin Calls: A New Driver of MMF Flows

### Key findings
- Investors into MMFs used MMFs to manage liquidity related to variation margin (VM) calls in the March 2020 market turmoil.
- VM payments faced by euro area non-bank financial sectors—in particular investment funds and pension funds—could have driven almost half of the aggregate outflows from EUR-denominated MMFs domiciled in the euro area.
- Investors may be reluctant to store liquidity received from VMs in MMFs during market stress or may postpone such decisions by a few days after receipt, reflecting high uncertainty.
- Results apply across all three countries in which EUR-denominated MMFs are domiciled in the sample: France, Ireland and Luxembourg.
- Analysis was conducted at country-sector-to-MMF level (not firm-to-MMF) because of data limitations.

### Quantitative estimates and mechanics
- German IFs are estimated to withdraw around €20 mn from each French MMF that they hold in order to post €1 bn of VM today (contemporaneous effect).
- German IFs are estimated to withdraw another €21 mn to post €1 bn of VM in two days (a 2-day forward estimate).
- Contemporaneous effects likely reflect the need to pay VM on centrally cleared trades; the 2-day forward estimate could reflect the need to pay VM on non-centrally cleared trades with less timely VM payments.
- Regarding inflows, German IFs are estimated to invest around €17 mn to €20 mn into each French MMF held by them one day after they received VM flow of €1 bn.
- On the same day of receiving VM flow of €1 bn, German IFs are estimated to invest around €4 mn to €5 mn into each Luxembourgish MMF held by them.
- The models with leads and lags tend to explain a slightly higher share of the actual aggregate MMF outflows and inflows compared with models without leads/lags.

### Model explanatory power (February–April 2020; regressions including leads/lags)
- MMF Outflows (one lead / two leads; Random Effects / Fixed Effects / Lagged Dep. Var.):
  - (1) All MMF outflows: 49% / 47% / 41%  and 47% / 44% / 40%
  - (2) Predict using significant coefs.: 49% / 46% / 44%  and 42% / 42% / 47%
  - (3) Only MMFs held by investors: 65% / 62% / 54%  and 62% / 58% / 53%
  - (4) Cases (2) and (3) combined: 65% / 61% / 59%  and 56% / 56% / 62%
- MMF Inflows (one lag / two lags; Random Effects / Fixed Effects / Lagged Dep. Var.):
  - (1) All MMF inflows: 28% / 14% / 37%  and 29% / 15% / 37%
  - (2) Predict using significant coefs.: 21% / 17% / 29%  and 9% / 10% / 12%
  - (3) Only MMFs held by investors: 38% / 19% / 50%  and 39% / 20% / 50%
  - (4) Cases (2) and (3) combined: 29% / 23% / 40%  and 13% / 13% / 16%

### Robustness and subperiods
- Results are robust to different model specifications (fixed effects, random effects, lagged dependent variable) and to inclusion of leads and lags of VM flows.
- Subperiod analysis comparing February–April and March shows variation in coefficient significance across countries and models (see regression tables for Luxembourgish, French, and Irish MMFs).

### Policy implications
- Non-bank liquidity management should account for the fact that the cash-like properties of MMF shares can fail under stress: MMFs can decline in value and can suspend redemptions in exceptional circumstances.
- MMFs should be made more resilient to significant outflows, and the structure of their investor base should be taken into account.
- Monitoring interconnectedness across markets and across borders is important, including from relatively small but volatile links introduced by daily margin exchange requirements.
- While daily exchange of high-quality collateral reduces counterparty credit risk, it can increase liquidity risk and create spillovers across markets.
- The unearthed link between VM payments and MMF flows suggests that macro- and market-stabilizing policy actions that reduce market volatility (and thus VM payments) can help stabilise MMF outflows.

### Data, limitations, and avenues for further research
- The analysis combines three highly granular unique datasets to construct a daily fund-level panel spanning from February to April 2020.
- Limitations: analysis could not be conducted at firm-to-MMF level due to data unavailability; results are therefore at country-sector-to-MMF level.
- Further research suggested on:
  - Non-bank liquidity responses beyond MMFs, including bank deposits, the repo market and credit or liquidity lines.
  - The role of initial margin payments and redemptions as additional liquidity drains on non-banks during stress.
  - Broader interconnectedness between the derivatives market and other markets using granular regulatory datasets.

*Source: Working Paper No. WP/2023/061*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2023/english/wpiea2023061-print-pdf.pdf_