An Attention-based Neural Network for Limit Order Books
Job Market Paper
Machine learning methods deliver superior forecasting accuracy but can hardly be used to make inference. To overcome this limitation, I propose an encoder-decoder neural network augmented with an attention-based mechanism that can autonomously learn to identify the most critical regions of the input data. I first train the model using high-frequency message data from the NASDAQ and show that it outperforms other state-of-the-art models in forecasting future transaction prices. Then, I develop a methodology that uses the attention mechanism to make inference on the relative share of information content of market orders versus limit orders, concluding that the most informative events are executions of market orders while submission and cancellations of limit orders are less relevant. Finally, I test the model's behavior during the execution of real block orders from institutional investors, showing that it favors liquidity provision rather than front-running strategies.
Evidence from Fire Sales
Published in the Journal of Finance
Using trade-level data, we study whether brokers play a role in spreading order flow information. We focus on large portfolio liquidations, which result in temporary drops in stock prices, and identify the brokers that intermediate these trades. We show that these brokers’ best clients tend to predate on the liquidating funds: at the beginning of the fire sale, they sell their holdings in the liquidated stocks, to then cover their positions once asset prices start recovering. The predatory trades generate at least 50 basis points over ten days and cause the liquidation costs for the distressed fund to almost double. These results suggest a role of brokers in fostering predatory behavior and raise a red flag for regulators. Moreover, our findings highlight the trade-off between slow execution and potential information leakage in the decision of optimal trading speed.
Evidence from the ETF Program of the Bank of Japan
Published in the Review of Asset Pricing Studies
Since the introduction of its Quantitative and Qualitative Easing program in 2013, the Bank of Japan has been increasing its holdings of Japanese equity through large scale purchases of index-linked ETFs with the intention of lowering assets' risk premia. We exploit the cross-sectional heterogeneity of the shock to supply induced by the policy to identify a positive, sizeable and persistent impact on stock prices consistent with a portfolio balance channel. The evidence suggests that demand curves for stocks are downward sloping in the long-run. We estimate an increase of 22 basis points in aggregate market valuation per trillion Yen invested into the program, corresponding to a price elasticity of 1. We show that the purchases of ETFs tracking the price-weighted Nikkei 225 index generate significant pricing distortions relative to a value-weighted benchmark. Finally, we provide a rigorous framework to discuss the consequences of a potential exit strategy from QE.
Asset Pricing Implications of Delta-Hedging
We build on a growing literature that studies the impact of market frictions on the dynamics of stock markets, such as momentum, price spirals, excess volatility, and investigate the potential feedback effects of delta-hedging in derivative markets on the underlying market. We document a link between large aggregate dealers’ gamma imbalances in illiquid markets and intraday momentum/reversal and market fragility. This link is distinct from information frictions (adverse selection and private information) and funding liquidity frictions (margin requirement shocks). We test our joint hypothesis using a large panel of index and equity options that we use to compute a proxy of aggregate gamma imbalance. We find supporting evidence that intra-day momentum (reversal) is explained by the interaction of negative (positive) aggregate gamma imbalance and market illiquidity. The effect is stronger for the least liquid underlying securities. The result helps to explain both intra-day volatility and autocorrelation of returns.
Factor Prediction in High Level Synthesis
Published in IEEE
High Level Synthesis development flows rely on user-defined directives to optimize the hardware implementation of digital circuits. Nevertheless, the most beneficial directive values are hard to predict, and exhaustive explorations are infeasible even for moderately complex designs. Focusing on the Loop Unrolling directive, we herein address this challenge by proposing a novel Machine Learning methodology, able to jointly forecast the optimal loop unrolling factors for all the loops in a target application. We showcase that our method results in a better prediction score (up to 63%) and a reduced convergence time compared to other state-of-the-art approaches. Our method achieves 90% of the speedup that can be obtained (with a perfect a-priori knowledge of optimal loop unrolling factors) when synthesizing the computational hotspots of each considered benchmark as hardware accelerators.