Amundi puts a Darwinian spin on bond portfolio rebalancing

‘Genetic’ algorithm picks bonds to buy or sell from quadrillions of possible combinations

Quants at Amundi Asset Management are using the principles of Darwinian natural selection to smooth the rebalancing of bond tracker portfolios.

The asset manager’s new ‘genetic’ algorithm helps portfolio managers convert subscription or redemption requests in corporate bond funds into workable trades within seconds, dramatically speeding up a task that previously could take several hours.

Passive funds track indexes containing thousands of bonds, each with minimum trade sizes and idiosyncratic liquidity considerations. Figuring out which bonds to buy or sell when investors come in and out of these funds is a heavy lift for even the most powerful computers.

Most firms rely on a form of dead reckoning when rebalancing, using algorithms that apply heuristics to estimate the closest practicable trades to theoretical best options.

Amundi’s algorithm mimics the evolutionary process in nature to reach what its developers claim is a better solution — and faster.

Genetic algorithms weigh hundreds or thousands of possible solutions to a given problem against one another, then combine features of the best and, through multiple iterations, find their way to a final outcome.

Investment managers have employed these types of algorithms elsewhere, such as to select which strategies to trade in a given market. 

For each rebalancing, Amundi’s algorithm speeds through 500 ‘generations’ of solutions, testing 5,000 specimen portfolios at each step. The best-performing portfolios ‘survive’ to populate the next phase. In each new iteration, some elements from the remaining portfolios are mixed and random ‘mutations’ added to create new candidate portfolios. “We have a solution in 20 or 30 seconds,” says Thierry Roncalli, head of quantitative research at Amundi. “It saves precious time for portfolio managers.”

The algorithm also makes backtesting of systematic strategies more accurate, Roncalli says.

In the coming weeks, Amundi will offer the tool through its Alto investment platform to clients. Roncalli’s team will also detail the workings of the algorithm in a paper to be published this week.

Trade balance

According to research firm ETFGI, there is more than $340 billion invested in exchange-traded funds (ETFs) tracking corporate bond indexes. Amundi started using the approach at the end of 2020 in some of its own corporate bond passive funds and ETFs, which have around €5 billion in assets.

If investors redeem €5 million from a fund tracking the Bloomberg Barclays Euro Corporate Bond Index, the fund manager would have to select just 30 bonds to sell from a universe of 3,000, Roncalli says. “You can have a big gap between the model portfolio and the investible portfolio.”

We have a solution in 20 or 30 seconds. It saves precious time for portfolio managers

Roncalli, Amundi Asset Management

A fund manager investing €500 million and up to €400,000 per bond in the index would have 243 quadrillion combinations to choose from. At 100 million portfolios a second, evaluating such choices fully would take 78 years.

Rebalancing transactions are worked out at present using heuristic algorithms with an element of human adjustment. Managers might carry out the task as often as five or six times a day for subscriptions and redemptions, as well as monthly to match index changes. Each transformation can take from five to 40 minutes, Roncalli says.

Raul Leote de Carvalho, deputy head of the quant research group at BNP Paribas Asset Management, says the firm’s multi-factor quant strategies have a similar issue, where liquidity constraints make it difficult to buy or sell the actual portfolio holdings immediately after subscriptions or redemptions.  

Genetic algorithms can be applied to solve such optimisation problems, he says, and in some cases can be more efficient. But he was unsure if a genetic algorithm would provide the fastest solution for the specific problem of multi-factor credit strategies his firm is grappling with.

The difficulty of selecting investible portfolios is a problem also when quants backtest systematic strategies, Roncalli says.

Such tests rely on model portfolios. But in real, life such portfolios would likely be impossible to trade. With the faster new algorithm, quants are able to backtest strategies using truly tradable portfolios over periods of 10 years or longer, Roncalli says.

Amundi recently started using the algorithm to backtest the effect of climate risk and environmental, social and governance (ESG) constraints on portfolios. The firm also plans to use its genetic algorithm in the design and backtesting of high-turnover corporate bond momentum strategies, where transaction costs can weigh heavily on returns.