Summary of a peer-reviewed paper
Banks can count some shares towards their liquidity buffer. Our research asked which shares actually hold up in a crisis, and found that the answer depends a lot on what kind of share it is.
Under the Liquidity Coverage Ratio, a bank may hold some equities in its liquidity buffer as Level 2B assets. They take a 50% haircut and can make up no more than 15% of the buffer. To qualify, a share must sit in a major index, and its price must not have fallen by more than 40% in a 30-day period of market stress.
Some argue that all of these equities should simply be treated as illiquid. That may be prudent, but it had not been tested against data. My co-author Murat Donduran and I wanted to know whether some shares behave better than others when markets seize up, and whether that can be predicted in advance.
We took 882 shares from the largest developed-market indices, with continuous data from January 2007 to June 2019. The test was the worst of the global financial crisis, September to November 2008. A share counted as illiquid if its cumulative price drop passed the 40% line the LCR uses. We then trained supervised machine learning models to predict that outcome from information known before the crisis: sector, industry, beta, market capitalisation and trading volume.
Most shares failed. 61% of them, 542 out of 882, fell by more than 40%. But the failures were far from even. Around four in five financial, basic materials and energy shares failed, against 37% in consumer non-cyclicals and only 24% in utilities.
Beta mattered too. Shares that held up had an average pre-crisis beta of 0.83, against 1.10 for those that did not, and higher beta went with illiquidity in every sector. Larger companies held up better in every sector except financials.
The best model, an ensemble using the RUSBoost algorithm, classified shares with a balanced accuracy of about 73%. Using a trading-based liquidity measure (Amihud) instead of price falls, it reached 82%.
An equity is not one asset class when it comes to liquidity. A more granular, risk-sensitive view of eligibility and haircuts is possible, and the data support it. Banks can use the same classification when they price funding and liquidity internally. As with any model, it needs controls, monitoring and expert judgement around it.
The study has limits. It covers developed markets and one main stress episode, so the next step is to test it on other crises and with a wider set of liquidity measures.
Liquidity classification of equities under stress using machine learning models: evidence from major world share indices