Objective research to aid investing decisions
Menu
Value Allocations for September 2019 (Final)
Cash TLT LQD SPY
Momentum Allocations for September 2019 (Final)
1st ETF 2nd ETF 3rd ETF

Predicting Stock Index Reversals with Amareos Sentiment Indicators

Posted in Sentiment Indicators

Do data-intensive, high-frequency investor sentiment measurements usefully predict stock index performance? In his May 2016 paper entitled “Can Sentiment Indicators Signal Market Reversals?”, Arnaud Lagarde applies a random forest machine learning algorithm to test the power of Amareos sentiment indications to predict stock index reversals. Algorithm training data relates sentiment to known stock index return for the next 182 days (six months). If this return is -20% or lower (+10% or higher), he designates the condition at the time of forecast as a market top (bottom). Otherwise, he designates the condition as neutral. He starts with 20 global equity indexes. He holds out four indexes (CAC40, CSI300, Nikkei and S&P500) for out-of-sample testing. He then randomly selects 80% of daily observations on the other 16 indexes for algorithm training, with the remaining 20% reserved for additional out-of-sample testing. Out-of-sample testing includes tabulation of raw top/bottom identification accuracy and a simple trading strategy that is long (in cash) after a bottom (top) indication and does not react to a neutral indications. He focuses trading strategy testing on: (1) the four hold-out indexes over the entire sample period; and, (2) the last six weeks of data for all indexes, which cannot be used for training. Using daily Amareos market sentiment readings and returns for the 20 equity indexes during January 2005 through mid-April 2016, he finds that:

  • Regarding accuracy of out-of-sample top/bottom signals:
    • Among all 20 indexes, 75.1% (92.4%) of bottom (top) signals are correct. However, 53.3% (59.6%) of bottom (top) signals are missed.
    • Among four hold-out indexes, 58.3% (37.2%) of bottom (top) signals are correct. However, 86.3% (96.5%) of bottom (tops) signals are missed.
  • Regarding trading strategy performance:
    • Based on both terminal value and maximum drawdown, the strategy outperforms buy-and-hold over the full sample period for all four hold-out indexes.
    • During the last six months of the sample period, the strategy outperforms buy-and-hold based on cumulative return for 11 of 20 indexes and matches buy-and-hold for 7 of 20. Average outperformance is 6.2%.

In summary, evidence from simple tests over a short sample period suggests that the Amareos sentiment signals may be useful for equity index timing.

Cautions regarding findings include:

  • The large day-to-day overlap (181 out of 182 days) in forecast intervals results in very high autocorrelation in forecasted returns and therefore probably much clustering of signals. Since sampling for training is random, it is very likely that training includes days from any given signal cluster, confounding discrimination between in-sample and out-of-sample data. Said differently, the sample period is very short in terms of independent 6-month return intervals, and the sampling methodology does not relieve this weakness.
  • Regarding section “3.3.2 Performances on CAC40, CSI300, Nikkei and S&P500”: The success of the strategy as applied to each of the four hold-out indexes derives essentially from avoiding one large drawdown (crash) during the sample period. Global equity indexes may have elevated return correlations during major crashes. Per the preceding caution, behaviors of the four hold-out indexes during crashes may be somewhat in-sample. [The author responds that “most of the ‘crashes’ are regional,” such as the CAC40 crash at the end of 2011 and the CSI300 crash at the end of 2015.]
  • Regarding section “3.3.3 Performance on the last six months”: As noted in the paper, this out-of-sample test is very short, with no changes in position for 7 of 20 indexes. During this subperiod, global equity markets generally decline. Even random market timing with a long-only strategy would tend to outperform buy-and-hold. [The author responds with a roughly estimated 27% probability that strategy outperformance is due to luck.]
  • As noted in the paper, strategy tests ignore trading frictions, which would lower performance. Per the paper: “However, because trading is limited – once or twice a year based on the 20 indices over the past decade – impact would definitely be minimal.” But, signals may tend to occur when liquidity is low (and trading frictions high). Also, the cost of the sentiment data reduces performance.
  • The methodology based on machine learning is likely beyond the reach of most investors, who would bear fees for delegating to an investment manager/fund.
Why not subscribe to our premium content?
It costs less than a single trading commission. Learn more here.
Daily Email Updates
Login
Research Categories
Recent Research
Popular Posts