Jupiter’s Systematic Equities team has explored the potential of using portfolio allocations and fund flows data to measure buy-side analyst sentiment on stocks. As a result of this project, they have managed to introduce a very useful diversifying factor into their alpha model, complementing the existing analyst sentiment stock selection criterion.
Extracting actionable data from the holdings of other fund managers has long attracted interest in academic literature, including analysis of breadth of ownership, concentration and fragility of holdings. However, an equally interesting area has attracted much less attention in the academic world: fund flows.
It stands to reason that if one knows how funds are performing, what they hold, and to what degree those funds may be in inflow or outflow, then one might divine some useful intelligence about the market impact of their trading activity. After all, persistence of mutual fund performances has been found to attract further fund flows, and such flows would typically be spent on existing portfolio holdings or trading stocks with similar characteristics. In the short term this forces prices to move in the same direction. The problem with this historically has been the availability of flow data.
Most academic studies are based on 13F filings, which concern mutual fund holdings, since flows can be derived from those in conjunction with published fund AUMs. A popular source collecting 13F data is Thomson Financial’s CDA/Spectrum Mutual Fund Holding Database, but its data is low frequency (quarterly), has limited coverage (mainly U.S. institutional), and has a long lag (45 calendar days). This means that it has little practical application for those seeking to gain an investment edge.
Views we can use
Those downsides have been overcome, however, by fund flow data from Emerging Portfolio Fund Research (EPFR), which is much timelier (monthly), with about 26 calendar days lags for holdings data and only 2 days lag for fund flows. The coverage of their dataset is also extended to global mutual funds and ETFs covering global equities. In addition to these two core sets of information, EPFR also provide more than a dozen quantitative factors based on one or both sets of information.
An important part of the EPFR dataset is portfolio allocations of fund managers, reflecting their views on the information for stocks held in their portfolios. Although EPFR is not allowed to divulge portfolio holdings of each individual fund, some measures of aggregated information (that cannot be used to reverse engineer individual fund managers holdings) are allowed and provide the basis for our investigations.
One way to utilise the information embedded in portfolio holdings is to look at all active fund managers, to form a picture of the aggregated positioning for all of them. Portfolios constructed to follow such allocations can lead to outperformance relative to market-weighted portfolios, which indicates that there is valuable information in the active fund manager allocations. Unavoidably, fund managers pick stocks based on certain themes, and such themes surface once groups of fund managers are aggregated.
Through analyses of the fund allocation data, it has been found that the active allocation signal has biases in a few common investment themes, for example, value, quality, price momentum and size. This is to be expected, since some or all of fund managers’ portfolio holdings are driven by one or some of those themes, sometimes intentionally and sometimes unintentionally. For our purposes, some forms of these common investment themes are already included in our alpha model, and therefore we must remove them to avoid double counting. After that signal purification, we find that there is still some excess alpha potential that is worth including in our model.
Following the money
The other part of EPFR data is for fund flows. Since money tends to chase fund performance, the flows into those performing funds would naturally lead the fund managers to add to their existing holdings and/or to trade stocks with similar characteristics. This in practice means generating positive momentum (or negative momentum in the case of fund outflows) for specific securities based on fund flows.
We have observed empirically that a fund flow momentum signal based on EPFR flow data has indeed demonstrated some usefulness in predicting future stock returns. Further analysis across GICS sectors, regions and countries also show that this is the case in most of the sub-categories. In addition to this, the flow momentum signal seems to decay within two to three months. This provided us with the opportunity to construct a signal to capture the information from fund flow. Due to the short duration and fast decay nature of the flow information, it makes sense to construct a short-term signal with more weighting given to the most recent data points. Once again existing biases for geographical regions, sectors, etc., also need to be accounted for.
Stronger together than apart
Conceptually, the above two signals represent two closely associated sets of information: allocation reflects aggregated buy-side analyst sentiment based on past information for the stock, whilst fund flows provide further up-to-date confirmation on whether such allocation is still in favour and investors are still putting money into it. Presumably, combining those two sets of information would be able to provide a more balanced and self-confirming forecast on a stock. Empirically, such combination does lead to a factor with diversification benefits. Various analyses we have conducted confirm that this composite flow factor (FLO) has a better return-risk profile than either of the two underlying signals individually across regions and across time.
We find the FLO factor to be orthogonal to existing alpha factors and its performance is broadly in line with them as well. The main impact of the FLO factor appears to be in diversifying the alpha model, and the performance of almost all our backtests that include FLO produce similar or slightly better returns with lower risk over the full history. Interestingly, weaker performance in the early period of the backtests is often more than offset by stronger performance since March 2018, leading to reduced drawdowns and improved Sharpe Ratios in particular. With such clear orthogonal information and diversification benefits, the FLO factor has been incorporated into our alpha model.
The value of Active Minds