In this article we look at few examples of the impact of adjusted vs. unadjusted data for dividends on features calculations used for an equity long/short strategy.
For the backtests in this article we used Norgate data for S&P 500 index that include current and past constituents to determine those stocks that have been in the index since January 2000. We also used DOW 30 weekly data for current constituents. We highly recommend this data service to users of DLPAL software (we do not have a referral arrangement with the company.)
We first determined that 215 S&P 500 stocks from the current composition have been in the index since January 2000 and exported two sets of data files for those stocks to use with DLPAL S software to calculate the features used in long/short strategies. The first set was split-adjusted data plus dividend adjustments and the second set just split-adjusted data.
The features for each S&P 500 stock were calculated daily for the period 01/02/2020 to 09/04/2020. Note that the objective of this study is not to develop a trading strategy but to study the relative impact of dividend adjustments on split-adjusted data. Therefore, neither the length of features series, nor the actual number of stocks used are that important as a first step; we just made sure that the data series start at the same time by selecting the 215 stocks that were in the index since January 2000.
Furthermore, the results of this study are limited, far from conclusive and just serve the purpose of further analysis. The issue of what type of data to use is important and differences in performance can be significant.
Strategies used in the study
We used two strategies as follows:
System with bias (not dollar neutral) – SYST
Maximum 10 positions, no minimum holding period.
Buy = Cover = Pdelta × S > 0
Short = Sell = Pdelta × S < 0
Rank with equal long and short (dollar neutral) – RANK
Maximum 5 long and 5 short positions, minimum holding period 2 days.
Score = Pdelta × S
The minimum holding period of two days is used to reduce the number of transactions since without it there is daily rebalancing. Pdelta is the directional bias and S is the significance of the bias, as calculated by DLPAL LS.
We used commission of $0.005 per share in all backtests below.
Case 1. S&P 500 Daily Data
The backtest period is from 01/02/2020 to 09/04/2020.
Below are the results of the backtest for SYST
| Parameter | Dividend Adjusted | Dividend Unadjusted |
| Return | -2.7% | +8.9% |
| Max. DD | -17.4% | -13.5% |
| Trades | 466 | 519 |
| Long trades | 266 | 299 |
| Short trades | 200 | 220 |
| Win% | 48.1% | 52.4% |
| Avg. holding time | 4.7 days | 4.3 days |
There is a large difference between adjusted and unadjusted data, with the latter type outperforming by a wide margin.
Below are the results of the backtest for RANK
| Parameter | Dividend Adjusted | Dividend Unadjusted |
| Return | -1.1% | +4.5% |
| Max. DD | -12.5% | -9.5% |
| Trades | 825 | 831 |
| Long trades | 415 | 416 |
| Short trades | 410 | 415 |
| Win% | 48.5% | 48.6% |
| Avg. holding time | 3.1 days | 3.1 days |
Also in this case the performance is considerably better in the case of unadjusted_for_dividends data. Note that dividend payouts should be included in the performance with unadjusted data and that would boost results somewhat. However, we did not go through this exercise due to its complexity and it was not required anyway since adjusted data performance was negative and unadjusted was positive.
Case 2. Dow 30 Weekly Data
The two strategies above were also used with weekly Dow 30 data. We used the current composition of the index for comparison purposes but we omitted V, DOW and CRM. For the 27 stocks considered, we exported weekly dividend adjusted and dividend unadjusted data since January 2000 and we instructed DLPAL LS to calculate weekly features starting on 01/05/2018 to 09/14/2020.
The difference with the S&P 500 strategies is that in the case of SYST, maximum 6 positions are allowed and in the case of RANK, maximum 3 long and maximum 3 short. Backtest period is features calculation period.
Below are the results of the backtest for SYST
| Parameter | Dividend Adjusted | Dividend Unadjusted |
| Return | +31.8% | +25.3% |
| Max. DD | -11.4% | -16.2% |
| Trades | 278 | 302 |
| Long trades | 157 | 165 |
| Short trades | 121 | 137 |
| Win% | 55.8% | 52.7% |
| Avg. holding time | 4.1 weeks | 3.8 weeks |
The differences between using dividend adjusted and unadjusted data could be probably attributed to dividend payouts in this case. There is no significant difference in the weekly timeframe due to data adjustments in the case of SYST and we actually expected that.
Below are the results of the backtest for RANK
| Parameter | Dividend Adjusted | Dividend Unadjusted |
| Return | 41.1% | +17.5% |
| Max. DD | -6.1% | -9.9% |
| Trades | 388 | 383 |
| Long trades | 197 | 193 |
| Short trades | 191 | 190 |
| Win% | 56.2% | 53.5% |
| Avg. holding time | 3.2 weeks | 3.2 weeks |
In this case, the deterioration in performance with unadjusted data cannot be explained by the lack of dividend payouts only and the adjusted series offer significant outperformance.
Conclusion
The choice between adjusted and unadjusted for dividend data for stocks (we always use split-adjusted to start with) is important and can lead to significant variations in backtested and actual performance. Our study in this article was basic and very limited. The study suggests that for daily timeframe it may be better to use unadjusted data to calculate the features and generate the signals while for weekly timeframe, the use of adjusted data could lead to better results. This is what we also suspected but this is a limited study and the results cannot be generalized. It is probably better to perform analysis for specific strategies using the two types of data series because performance may depend also on the type of strategy.
For more information about DLPAL LS click here.
The program manual can be found here.
We offer daily S&P 500 long/short signals to professional traders and hedge funds. Click here for more details.
