Universal features of price formation in financial markets: perspectives from deep learning

Using a large-scale Deep Learning approach applied to a high-frequency database containing billions of market quotes and transactions for US equities, we uncover nonparametric evidence for the existence of a universal and stationary relation between order flow history and the direction of price moves. The universal price formation model exhibits a remarkably stable out-of-sample accuracy across a wide range of stocks and time periods. Interestingly, these results also hold for stocks which are not part of the training sample, showing that the relations captured by the model are universal and not asset-specific.

The universal model—trained on data from all stocks—outperforms asset-specific models trained on time series of any given stock. This weighs in favor of pooling together financial data from various stocks, rather than designing asset- or sector-specific models, as is currently commonly done. Standard data normalizations based on volatility, price level or average spread, or partitioning the training data into sectors or categories such as large/small tick stocks, do not improve training results. On the other hand, inclusion of price and order flow history over many past observations improves forecast accuracy, indicating that there is path-dependence in price dynamics.     

Neuro‐fuzzy time‐series analysis of large‐volume data

This paper describes a framework that utilizes an adaptive‐network‐based fuzzy inference system to perform user‐constrained pattern recognition on time‐series data. Using a customizable fuzzy logic grammar, the architecture allows an analyst to capture domain expertise in a context‐relevant manner. Fuzzy logic rules constructed by the analyst are used to perform feature extraction and influence the training of a neural network to perform pattern recognition. We demonstrate that the architecture is capable of performing noise‐tolerant searches across multiple features on large volumes of time‐series data. The experiments presented here are from the domain of stock analysis. We are able to create simple rule sets automatically to search a data warehouse of stocks to select stocks that exhibit desirable behaviours. Copyright © 2011 John Wiley & Sons, Ltd.     

Commonalities in the order book

This paper uses data from one of the most important European stock markets and shows that, in line with predictions from theoretical market microstructure, a small number of latent factors captures most of the variation in stock specific order books. We show that these order book commonalities are much stronger than liquidity commonality across stocks. The result that bid and ask side as well as the visible and hidden parts of the order book exhibit quite specific dynamics is interpreted as evidence that open order book markets attract a heterogeneous trader population in terms of asset valuations and impatience. Quantifying the informational content of the extracted factors with respect to the evolution of the asset price, we find that the factor information shares are highest (about 10%) for less frequently traded stocks. We also show that the informational content of hidden orders is limited.

Encoding of high-frequency order information and prediction of short-term stock price by deep learning

Predicting the price trends of stocks based on deep learning and high-frequency data has been studied intensively in recent years. Especially, the limit order book which describes supply-demand balance of a market is used as the feature of a neural network; however these methods do not utilize the properties of market orders. On the other hand, the order-encoding method of our prior work can take advantage of these properties. In this paper, we apply some types of convolutional neural network architectures to order-based features to predict the direction of mid-price trends. The results show that smoothing filters which we propose to employ rather than embedding features of orders improve accuracy. Furthermore, inspection of the embedding layer and investment simulation are conducted to demonstrate the practicality and effectiveness of our model.     

New Paradigms in Stock Market Indexing

Considerable recent interest has been shown in a new set of stock‐market indices that are weighted by fundamental factors such as sales, earnings, dividends or book values, rather than by capitalization. In this paper, we analyze the performance of Fundamental Indexing? (“FI”). First, we show that the source of FI's recent excellent performance is not from its ability to systematically arbitrage mispricing in a noisy market but from increasing the portfolio's exposure to stocks with low price‐to‐book values and with small capitalizations. We find that FI does not produce a positive alpha when its excess returns are explained by the Fama‐French three‐factor model of CAPM beta, the value premium and the size premium. Second, we show that it is possible to construct a portfolio of exchange‐traded funds with similar factor loadings that can replicate, and sometimes, even outperform FI. However, we caution investors not to expect consistent outperformance from portfolios tilted towards value and small‐cap stocks. Historical data shows evidence of mean reversion in the performance of such strategies.     

Decomposing liquidity along the limit order book

In this paper, we estimate the asymmetric information and order processing components of liquidity at extended depths along the limit order book. Using data from the INET ECN, we find that the asymmetric information component decreases as depth increases. Inactive stocks have more information asymmetry at the inside quotes, but it decreases more rapidly along the book than for active stocks. The order processing component of liquidity has fixed and variable constituents, consistent with a fixed cost per order as well as a variable portion based on the size of the order.     

Style‐Related Comovement: Fundamentals or Labels?

I find that economically meaningless index labels cause stock returns to covary in excess of fundamentals. S&P/Barra follow a simple mechanical procedure to define their Value and Growth indices. In doing so, they reclassify some stocks from Value to Growth even after their book‐to‐market ratios have risen, and vice versa. Such stocks begin to covary more with the index they join and less with the index they leave. Backdated constituent data from Barra reveal no such label‐related shifts in comovement during the 10 years prior to the actual introduction of the indices in 1992.     

The effect of organizational structure on efficiency: Evidence from the Spanish insurance industry

This paper provides new information on the effects of organizational structure on efficiency by analyzing Spanish stock and mutual insurers over the period 1989–1997. We test the efficient structure hypothesis, which predicts that the market will sort organizational forms into market segments where they have comparative advantages, and the expense preference hypothesis, which predicts that mutuals will be less efficient than stocks. Technical, cost, and revenue frontiers are estimated using data envelopment analysis. The results indicate that stocks and mutuals are operating on separate production, cost, and revenue frontiers and thus represent distinct technologies. In cost and revenue efficiency, stocks of all sizes dominate mutuals in the production of stock output vectors, and smaller mutuals dominate stocks in the production of mutual output vectors. Larger mutuals are neither dominated by nor dominant over stocks in the cost and revenue comparisons. Thus, large mutuals appear to be vulnerable to competition from stock insurers in Spain. Overall, the results are consistent with the efficient structure hypothesis but are generally not consistent with the expense preference hypothesis.     

Average Returns,B/M,and Share Issues

The book‐to‐market ratio (B/M) is a noisy measure of expected stock returns because it also varies with expected cashflows. Our hypothesis is that the evolution of B/M, in terms of past changes in book equity and price, contains independent information about expected cashflows that can be used to improve estimates of expected returns. The tests support this hypothesis, with results that are largely but not entirely similar for Microcap stocks (below the 20^th NYSE market capitalization percentile) and All but Micro stocks (ABM).     

Fails‐to‐Deliver before and after the Implementation of Rule 203 and Rule 204

We study the determinants of fails‐to‐deliver in the period before and after the implementation of Rule 203 (elimination of option market maker exception from the locate and close‐out requirement) and Rule 204 (t+3 close‐out rule) in September 2008. We find a positive relation between short selling and fails‐to‐deliver that weakens after the implementation of these rules. Fails‐to‐deliver are higher for stocks with low institutional ownership, low book‐to‐market, small market capitalization, high turnover, and put option availability. The relation between short selling and these measures of borrowing costs is also weaker after the implementation of these rules.     

The Cross-Section of Volatility and Expected Returns

We examine the pricing of aggregate volatility risk in the cross‐section of stock returns. Consistent with theory, we find that stocks with high sensitivities to innovations in aggregate volatility have low average returns. Stocks with high idiosyncratic volatility relative to the Fama and French (1993, Journal of Financial Economics 25, 2349) model have abysmally low average returns. This phenomenon cannot be explained by exposure to aggregate volatility risk. Size, book‐to‐market, momentum, and liquidity effects cannot account for either the low average returns earned by stocks with high exposure to systematic volatility risk or for the low average returns of stocks with high idiosyncratic volatility.     

Contrarian Profits and the Overreaction Hypothesis: the Case of the Athens Stock Exchange

This paper investigates the existence of contrarian profits and their sources for the Athens Stock Exchange (ASE). The empirical analysis decomposes contrarian profits to sources due to common factor reactions, overreaction to firm‐specific information, and profits not related to the previous two terms, as suggested by Jegadeesh and Titman (1995). Furthermore, in view of recent evidence that common stock returns are related to firm characteristics such as size and book‐to‐market equity, the paper decomposes contrarian profits to sources due to factors derived from the Fama and French (1993, 1996) three‐factor model. For the empirical testing, size‐sorted sub‐samples that are rebalanced annually are employed, and in addition, adjustments for thin and infrequent trading are made to the data. The results indicate that serial correlation is present in equity returns and that it leads to significant short‐run contrarian profits that persist even after we adjust for market frictions. Consistent with findings for the US market, contrarian profits decline as one moves from small stocks to large stocks, but only when market frictions are considered. Furthermore, the contribution to contrarian profits due to the overreaction to the firm‐specific component appears larger than the underreaction to the common factors.     

Do Cash Flows of Growth Stocks Really Grow Faster?

Contrary to conventional wisdom, growth stocks (i.e., low book‐to‐market stocks) do not have substantially higher future cash‐flow growth rates than value stocks, in both rebalanced and buy‐and‐hold portfolios. Efficiency growth, survivorship and look‐back biases, and the rebalancing effect help explain the results. These findings suggest that duration alone is unlikely to explain the value premium.     

Interaction of size,book‐to‐market and momentum effects in Australia

This study seeks to disentangle the effects of size, book‐to‐market and momentum on returns. Initial results show that each characteristic has a role in explaining returns, but that there is interaction between size and momentum, as well as between size and book‐to‐market. Three key findings emerge. First, the size premium is the strongest, particularly in the loser portfolios. Second, the value premium is generally limited to the smallest portfolios. Third, the momentum premium is evident for the large‐ and middle‐sized portfolios, but loser stocks significantly outperform winner stocks in the smallest size portfolio. When these interactions are controlled with multivariate regression, we find a significant negative average relation between size and returns, a significant positive average relation between book‐to‐market and returns, and a significant positive average relation between momentum and returns.     

Expectations and the Cross-Section of Stock Returns

Previous research has shown that stocks with low prices relative to book value, cash flow, earnings, or dividends (that is, value stocks) earn high returns. Value stocks may earn high returns because they are more risky. Alternatively, systematic errors in expectations may explain the high returns earned by value stocks. I test for the existence of systematic errors using survey data on forecasts by stock market analysts. I show that investment strategies that seek to exploit errors in analysts' forecasts earn superior returns because expectations about future growth in earnings are too extreme.     

Housing Collateral, Consumption Insurance, and Risk Premia: An Empirical Perspective

In a model with housing collateral, the ratio of housing wealth to human wealth shifts the conditional distribution of asset prices and consumption growth. A decrease in house prices reduces the collateral value of housing, increases household exposure to idiosyncratic risk, and increases the conditional market price of risk. Using aggregate data for the United States, we find that a decrease in the ratio of housing wealth to human wealth predicts higher returns on stocks. Conditional on this ratio, the covariance of returns with aggregate risk factors explains 80% of the cross‐sectional variation in annual size and book‐to‐market portfolio returns.     

Analyst Forecasts and the Cross Section of European Stock Returns

We examine revisions to earnings forecasts by equity analysts and their role in predicting stock returns. We provide evidence that European stocks with net upward revised forecasts earn higher future returns than otherwise similar stocks. This effect is not concentrated in small stocks, stocks with low analyst coverage, or stocks with low book‐to‐market ratios. We find differences in the return continuation patterns of stocks with upward versus downward revisions, namely, bad news travels quickly, but good news travels slowly. This result is consistent with investors' attaching greater significance to poor earnings forecasts, but adopting a wait‐and‐see approach to good news.     

Is Timing Everything? The Value of Mutual Fund Manager Trades

I develop new measures of the value of active mutual fund management using portfolio holdings. These measures simultaneously test for trading and selection skill within stocks, industries, and characteristics. I demonstrate that most of the skill documented in prior studies comes from correctly trading stocks within industries, though funds also have some skill in timing industries. However, prior research focuses on the period 1980‐1994. I also test the hold out sample 1995‐2007. Contrary to prior results, the latter period (and the full sample) demonstrates that mutual funds generate no excess returns from any category of skill.     

Banning short sales and market quality: The UK’s experience

We study the effects that the ban on short sales of shares in financial firms introduced in late 2008 and removed early 2009 had on the microstructure and the quality of UK equity markets. We show that the ban did nothing to affect order flows: financial stocks were being more aggressively sold off than their peers pre-ban and this situation persisted through the ban period. Trading volume in financials was massively reduced, however. The ban decimated order book liquidity for financials. The deterioration was symmetric, affecting the limit buy and limit sell side of the order book equally. Finally we show that, through the period of the ban, markets for financial stocks were substantially less efficient and that the role of the trading process in aiding price discovery was greatly reduced. The effects identified above were largely reversed once the ban was lifted. The persistence of the deterioration in market quality and liquidity though the relatively long-lasting UK ban on short selling suggests that other major market developments such as the TARP program were not responsible since these were concentrated in the early half of the ban. We thus argue that the short selling ban was responsible for detrimental effects on the quality of UK equity markets and that, far from being stabilising, the ban exacerbated problems in valuing UK financial stocks.     

Information, Trading, and Product Market Interactions: Cross-sectional Implications of Informed Trading

I present a simple model of informed trading in which asset values are derived from imperfectly competitive product markets and private information events occur at individual firms. The model predicts that informed traders may have incentives to make information‐based trades in the stocks of competitors, especially when events occur at firms with large market shares. In the context of 759 earnings announcements, I use intraday transactions data to test the hypothesis that net order flow and returns in the stocks of nonannouncing competitors have information content for announcing firms.