212 Years of Price Momentum (The World’s Longest Backtest: 1801-2012) by SSRN
University of Pennsylvania – The Wharton School, Finance Department
Forefront Analytics
August 1, 2013
Abstract:
We assemble a dataset of U.S. security prices between 1801 and 1926 and create an out-of-sample test of the price momentum strategy, discovered in the post-1927 data. The pre-1927 momentum profits remain positive and statistically significant. Additional time series data strengthen the evidence that momentum is dynamically exposed to market beta, conditional on the sign and duration of the tailing market state. In the beginning of each market state, momentum’s beta is opposite from the new market direction, generating a negative contribution to momentum profits around market turning points. A dynamically hedged momentum strategy significantly outperforms the un-hedged strategy.
212 Years of Price Momentum (The World’s Longest Backtest: 1801-2012) – Introduction
The first two U.S. stocks traded in 1792 in New York. Over the following decades, the securities market developed rapidly. By the end of 1810, 72 traded securities existed, and by the end of the 1830s the number was more than 300. To our knowledge, all current academic studies of U.S. security-level data begin in 1926, the year the CRSP database began. The U.S. market had been active for 133 years before that time, providing an opportunity to test stock-level studies in earlier history. The 19th and early 20th centuries are filled with expansions, recessions, wars, panics, manias, and crashes, all providing a rich out-of-sample history. Limiting studies to the post-1925 period introduces a strong selection bias and does not capture the full distribution of possible outcomes.
For example, the case of price momentum: Before 2009, only following the Great Depression did the strategy have a decade-long negative compounded return. Such an occurrence was considered to be an outlier and the remaining part of the distribution was understood as normal. Since 2009, coming out of the second-worst U.S. financial collapse, momentum has experienced another decade-long underperformance, creating a large ripple in investment portfolios that use this strategy. The repeated underperformance raised practical questions about the outlier conclusion and what the actual distribution of momentum profits is. By extending the momentum data back to 1801, we create a more complete picture of the potential outcomes of momentum profits, discovering seven additional negative decade-long periods prior to 1925.
The first contribution of this study is a creation of a monthly stock price dataset. In this dataset, three known 19th and early 20th century data sources are combined into one testable dataset from 1800 to 1927. Those data sources are from the International Center of Finance at Yale (ICF), the Inter-University Consortium for Political and Social Research (ICPSR), and Global Financial Data (GFD). Between 1800 and 1927, the merged dataset contains an average of 272 securities per month, making it robust for security-level studies.
The second contribution of this study is to add to the existing price momentum literature by extending the momentum tests to the new data. Our study finds that in the pre-1927 data, the momentum effect remains statistically significant and is about half that of the post-1927 period. From 1801 to 1926, the equally weighted top third of stocks sorted on price momentum outperformed the bottom third by 0.28% per month (t-stat 2.7), compared to 0.58% per month (t-stat 3.6) for the 1927-2012 period. Linking the two periods together generates a 212-year history of momentum returns, averaging 0.4% per month (t-stat 5.7).
As observed in the studies of the 20th century data, momentum profits are highly variable over time, giving rise to the limits-of-arbitrage explanation. Nevertheless, over the long run, the trend-following strategy would have generated significant market outperformance, in a different century than the one in which it was discovered and tested. Our study adds to the evidence that momentum effect is not a product of data-mining but is highly variable overtime.
The third contribution of this study is to link momentum’s beta exposure to the market state duration. We find strong evidence that momentum beta is positively exposed to the duration of both positive and negative market states. The longer a given market state persists, the stronger the momentum portfolio beta exposure becomes. Analyzing the longer history is especially useful for the time-series tests, as the sample size is more than doubled.
See full PDF below.