The machine shines in Hong Kong A-share fund

Strategy run by ChinaAMC (HK) combines machine learning with human judgement to outdo rivals

Don Huang of Magnum Research (left) and Frederick Chu of ChinaAMC (HK)

China’s falling stocks have wrong-footed scores of fund managers this year but a Hong Kong-based quantitative fund has a different story to tell.

The multi-factor strategy, long-only fund, managed by China Asset Management (Hong Kong) or ChinaAMC (HK), invests in yuan-denominated equities of China’s mainland companies known as A-shares. It has returned 6% so far this year, while peers have lost money.

The fund’s so-called AI China Alpha Strategy uses a machine learning algorithm to identify pricing patterns in market data and pick stocks based on a constantly evolving mix of factors. It regularly applies human judgement to “polish” the algorithm, focuses on a relatively small number of the most liquid stocks and aims to avoid market impact.

“The inefficiency in the A-shares market makes it an ideal ground for a systematic portfolio approach,” says Frederick Chu, who is in charge of exchange-traded funds at ChinaAMC (HK) – a subsidiary of one of China’s largest asset managers, China Asset Management Co.

The inefficiency stems from dominance of the market by domestic retail investors: they make up around 80% of trading on China’s mainland, which itself accounts for the bulk of activity in A-shares. As retail investors in what are still young capital markets are less sophisticated, the result is frequent mispricing of stocks.

“We were quite determined that data would uncover a number of behavioural patterns within the A-shares market,” Chu says.

So, last year, Chu and his team approached a Hong Kong-based financial technology start-up Magnum Research to come up with an algorithm for the strategy.

A 10-strong team at the start-up selected variables from datasets – including details on companies, macroeconomic and share pricing data and information about investor behaviour – that they expected to be predictive of factor performance.

Using these ‘features’, the algorithm then identifies and continuously adjusts the combination of factors that works best in a certain market environment. This contrasts with traditional factor investing products, which are bound by a static set of factors.

One discovery the statistical model has made is that retail investors tend to lag institutional investors in responding to positive signals about a company.

“Historic data shows that there is a short-term momentum pattern where price is further lifted by the retail group as positive information is released,” says Don Huang, the co-designer of the start-up’s Aqumon investment engine, which powers the strategy. “Over a longer term, we’ve noticed the price would peak after a large number of retail investors enter and [then] the stock price slips back to a reasonable valuation.”

Human touch

But Huang’s team also keeps the algorithm up-to-date. It monitors structural changes affecting the A-share market – such as the stocks’ inclusion in coveted MSCI indexes and China’s efforts to further open up its capital markets – and evaluates the accuracy of the model’s predictions on a regular basis. This is the strategy’s human element.

“At present, the technology of deep learning is yet to reach a completely autonomous level once the algorithm is developed. Human understanding is still needed from time to time to polish the algorithm in order to ensure its relevance,” Huang says.

Another task that requires human intervention is adjusting the price of a stock in the fund after its trading has been suspended – something that afflicts China’s onshore shares particularly often, preventing investors from exiting positions. The practice reached a peak during the Chinese stock market turmoil that started in June 2015: 51% of companies that make up the A-share market had their shares suspended that month.

“A fair valuation model is created to assess the likely impact on stock prices after every suspension notice is published,” says Chu. Analysts at his firm use the model, combined with some judgement, to adjust the valuations.   

ChinaAMC (HK) and Magnum Research have run the strategy since the beginning of 2018, starting before its official launch in March. So far this year, the factor most favoured by the algorithm has been quality and it has contributed the most of the strategy’s outperformance, Huang says.

Although the computer can perform some analysis beyond human cognitive ability, the algorithm cannot completely avoid underperformance

Don Huang, Magnum Research

As well as maximising returns, the strategy is designed to minimise risks.

When rebalancing each month, the model also produces a score for the intensity of market signals, which Huang likens to a human manager’s confidence level in certain positioning. Strong market signals typically result in concentration in fewer stocks, whereas weaker market patterns lead to greater diversification in the portfolio. Such adjustments happen automatically but the model displays the ‘confidence’ score so that managers can see the rationale. 

The strategy invests in a limited number of equities – between 30 and 60 at different points in time – selecting from around 1,400 A-shares that are open both to onshore and offshore investors via China’s cross-border Stock Connect scheme. This approach ensures “liquidity and tradability of stocks”, Huang says.

The model also puts limits on the daily trading volume of each stock to avoid market impact.

For now, ChinaAMC (HK) is promoting the fund only to a small number of private and institutional investors. In particular, the firm is looking to attract high-net-worth individuals based in Hong Kong, who tend to be sophisticated enough to comprehend the mechanism, Chu says.

Chu and Huang emphasise to clients that the computer picks stocks based on probability, rather than certainty, of outperformance.

“The process is disciplined,” Huang says. “[But] although the computer can perform some analysis beyond human cognitive ability, the algorithm cannot completely avoid underperformance.”

So far, though, the machine has excelled. 

In stark contrast with the fund’s success this year, Hong Kong-based peers with a similar investment mandate have lost a fourth of their value, according to data from Morningstar Direct. The A-share market, meanwhile, is on course for its worst annual performance in a decade, ravaged by a trade war with the US, China’s slowing economy and tightening global liquidity.

Editing by Olesya Dmitracova