Sell, Resell: The Unexpected Election of Donald Trump in 2016 and the Behavior of the New York Stock Exchange
In this research paper, we utilize the unexpected victory of Trump in the 2016 American election to investigate the behavior of the New York Stock Exchange. The unexpected nature of the event provides an interesting research opportunity on political events and their influence on stock market behavior. How did the New York Stock Exchange (NYSE) react to the unexpected election of Donald J. Trump in 2016? We expect that the surprising victory immediately altered the stock exchange’s trajectory. Our main hypothesis is that on average, the daily high of the NYSE significantly differs in the periods before and after the event of Trump’s unexpected victory.
Causal Effect and Assumptions
We formalize the expected causal relationship in the “possible outcomes” framework. Our dependent variable is the daily high of the NYSE for each day from October until December of 2016. Our binary treatment variable is whether the trading day in question preceded or proceeded the election date. 1 represents days after the 8th of November 2016, and 0 before that date. As such, the election day itself, November 8th, does not fall into the treatment category. Our causal treatment effect is the average difference in our dependent variable, the daily high of the NYSE, between the days in the pre-election and post-election period.
We assume that the treatment and control group only differ in the sense that traders have opposing expectations of the election results. In the counterfactual outcome, the average stockbroker would have expected Trump to win. Since this is not observable to make viable ‘all else equal’ comparisons, we assume that the pre-election period closely represents our counterfactual outcome. We assume all other characteristics of the days in both the control and treatment period to be randomly distributed. Our control variables will seek to control for this assumption (see Table 1, Model 2-4).
Operationalization
Our chosen data for the dependent variable are the observed daily highs of the NYSE, retrieved from the database Kaggle (2008-2023). The term Daily high translates to the highest intraday price for a security on a given trading day (Cheung, 2007). This unit of measurement was chosen as a macroeconomic measurement of day-to-day outlook on the American market. Furthermore, daily high is a measurement of great importance to economic development and future investment potentials. This outcome variable has been transformed into a baseline-adjusted variable, to simplify the numbers from 5-digits to a percentage value. The mean value of the entire observational period of 63 days, from October 3rd – December 30th was taken as 100%, referencing all other observations relative to said value. The mean of the dependent variable is 10814.36 points (100%), the index ranges from 10352.98 (95.7%) points to 11688.45 (108.08%) points, with a standard deviation of 273.71 (2.53 %).
Figure 1. NYSE Daily high, October 3rd-December 30th, 2016.
The treatment variable is a dummy variable classifying days in the pre-election period (October 3rd-November 8th) as 0 while classifying every day in the post-election period (November 9th-December 30th) as 1. This way, we can discern if any given day was before or after our unexpected event.
Covariates
We assume that the level of precipitation may influence the daily high of the NYSE. Stockbrokers may find sunnier days lucrative for investments, thereby confounding the initial hypothesis H1. In the analysis, we thereby include daily average precipitation from Central Park, New York City for the two time periods. The variable is then transformed into a binary dummy variable in order to more easily categorize possible relations. 0 means no rain on the given day, 1 means that it rained or snowed in any capacity. We further assume that oil prices may influence the daily high of the NYSE. As an independent source of influence on stock markets and the economy in general, the importance of oil should not be understated. A new, unexpected president that might implement lenient fossil fuel policies could alter expected crude oil sales, which drives stock market performance. For this reason, we include an ordinal variable of Oil price futures as a mediating variable, meaning it either deteriorates or enhances the effect of our pre-post treatment on our dependent variable.
The scale of the Oil price futures variable is a three-point ordinal variable. It takes the value 1 if the price of the day in question is within the first quantile (<= 47.8 USD) of the prices’ distribution in the observation period (3rd October – 30thof December), 2 if the price of the day is within either the second or third quantile (> 47.8 USD & <= 51.8 USD), and 3 if the price of the day is in the fourth and final quantile (> 51.8 USD) of the distribution.
Results & Interpretation
Our analysis proceeds in four different regression models. Model 1 will be a bivariate modelling of the relationship between our dependent variable and the pre-post-election day treatment. Model 2 and 3 will each build on model 1, adding the Rain control variable, and the Oil price futures mediating variable respectively in isolation. In model 4 we include all the variables in combination. Each model uses robust standard errors. The results for model 1-4 are presented in Table 1, with results rounded up to four digits. In model 3 and 4 two observations were omitted due to missing values in the Oil price futures variable, reducing the number of observations to 61 days from 63.
Table I. Results of Regression Models (baseline-adjusted)
| Model 1 | Model 2 | Model 3 | Model 4 |
Intercept | 0,9783 (0,0018) *** | 0,9787 (0,0018) *** | 0,945 (0,0039) *** | 0,945 (0,0038) *** |
‘Pre-Post Treatment’ | 0,0379 (0,0039) *** | 0,0392 (0,0046) *** | 0,0337 (0,0024) *** | 0,035 (0,0032) *** |
Rain | - | -0,0051 (0,0052) [0.333] | - | -0,0062 (0,0032) [0.058] |
Oil Price Futures | - | - | 0,0181 (0,0021) *** | 0,0182 (0,0020) *** |
Adjusted R2 | 0,5578 | 0,5643 | 0,8309 | 0,8404 |
N | 63 | 63 | 61 | 61 |
*** p≤.001, **p≤.01, *p≤.05. Standard errors within parentheses. Source: Kaggle (Pavan Narne), Global Stock Market (2008-2023). URL: https://www.kaggle.com/datasets/pavankrishnanarne/global-stock-market-2008-present?select=2016_Global_Markets_Data.csv Comment: unstandardized b coefficients, standard errors in parentheses.
Figure 2: Coefplot of Model 1-4
Model 1 already shows promising results for our main hypothesis that the unexpected nature of the 2016 election did in fact alter the daily performance of the NYSE. With an intercept of 0.9783 and an effect size of 0.0379 on our treatment variable, we can predict that a day in the treatment group on average will put the daily high of the NYSE above the mean value in the observation period. This effect is highly significant with a p-value below 0,001 and a confidence interval ranging from 0,0301 to 0,0457.
Model 2 confirms the significant effect of our treatment variable on the outcome variable, with marginal changes in effect and significance. However, we do not observe a significant effect of the Rain covariate, with a sub-zero t-value and a p-value of 0,333. The effect size of Rain on the outcome variable itself is negligible with a change in value of -0.0051 on days where it did not rain. In addition, with a confidence interval going from –0,0156 to 0,0054, we cannot reliably estimate that it is not a non-zero effect. This relationship does not change in model 4 once all covariates are considered in conjunction, with a final p-value of the Rain covariate in model 4 of 0,058.
Model 3 displays an interesting relationship of the covariates. When the Oil price futures covariate is introduced, the effect of our treatment decreased, which aligns with our expectations as oil prices are known to also be related to stock market performance. Being of positive nature (0,0182), the effect of high oil price futures does seem to boost the pre-post treatment effect. The treatment’s size decreases to 0,0337, but the adjusted R2 increases from 0,5578 to 0,8309 indicating fewer residuals between the data and the fitted model. Moreover, adding this additional coefficient does not alter the linear relationship and the significance found in model 1 in any meaningful way. Thus, controlling for the daily oil price in our observation period, confirms our hypothesis while gaining a more nuanced picture of influences on the daily high of the NYSE.
In Model 4, where all variables are considered in conjunction, we see robust results between our dependent variable and the covariates maintained. The adjusted R2 is 0,8404 and is the highest among the four models, indicating a well fit of the linear model. Reaching a first conclusion, the covariate Rain can safely be ousted as an important factor in the analyzed relationship between the unexpected victory of Trump and NYSE performance in the pre- and post-election periods.
Predictions
Based on model 4, we can make valid prediction of NYSE behavior utilizing the coefficients present in the formula above. Consider this first example (result calculated using unrounded numbers):
We interpret the result as the estimated baseline-adjusted daily high of a day before the election, when it rained, and when the oil price futures characteristic was in the lowest quantile (<= 47,8 USD). On these days, we estimate that on average, the daily high was performing 4,2526 percentage points below the baseline. In contrast, consider this second example:
Outliers & Violations
December 27th, the first open stock market day after Christmas Day in 2016, is an outlier in terms of daily high, with a much larger value than other days in the post-period. While this is explained through the concept of a holiday boom (Groette, 2024), in which stocks tend to give a higher return on investment just after holidays, it may pose a problematic violation in terms of validity to our average treatment effect. For this reason, we chose to replicate model 4 with December 27th omitted from the post-period and review the results in comparison to the initial model.
Table 2: Adjusted Model 4, Dec 27th omitted | |
Intercept | 0,948 (0,0028)*** |
‘Pre-Post Treatment’ | 0,0339 (0,0023)*** |
Rain | -0.0045 (0.0027) [0,103] |
Oil Price Futures | 0.0167 (0.0014)*** |
Adjusted R2 | 0,8759 |
N | 60 |
*** p≤.001, **p≤.01, *p≤.05. Standard errors within parentheses.
Source: Kaggle (Pavan Narne), Global Stock Market (2008-2023). URL: https://www.kaggle.com/datasets/pavankrishnanarne/global-stock-market-2008-present?select=2016_Global_Markets_Data.csv Comment: unstandardized b coefficients,
standard errors in parentheses.
With the number of observations reduced to 60, we see expected changes in the results. The intercept increases slightly, with a minor increase to the standard error. The pre-post and oil prices also decrease in effect size, again in line with expectations. The main observation is that with the outlier omitted, the overall fitness of the model improves. This can be observed by looking at the adjusted R2, going from 0,8404 to 0,8759. Notably, the Rain covariate becomes even less significant, with a p-value increase from 0,058 to 0,103.
An additional violation of our model pertains to autocorrelation. Stock security could have a strong positive autocorrelation of returns, since a price increase on any given day suggests an increased chance of further price increase the day after (Cheung, 2007). This may skew the predictability of the models. To overcome this violation of OLS assumptions we may use a lagged variable of our dependent variable to further isolate the observational period from overall time series effects. Other than that, we believe our model is a good representation of real causal mechanisms. We can make accurate and significant predictions of the relationships between the outcome variable and covariates used.
Conclusion
In summary, we find support for our relationship between the pre-post periods and its effect on the NYSE daily high. We were able to confirm our hypothesis that the daily high of the NYSE significantly differs in the periods before and after the event of Trump’s unexpected victory. We find mixed results for our controlling variables, with Rain not having any significant effect. On the other hand, Oil price futures positively mediate the treatment effect, although causality is not settled conclusively. Omitting the outlier was a necessary step, since all correlations increased in significance while doing so. Continuing the analysis with a lagged dependent variable may further improve the model’s predictability and validity, while also accounting for a larger timeframe and other possible events within it. Moreover, applying a synthetic control research design on the question of presidential campaigning and its effects on macroeconomic performance prospects aides the generalizability of the hypothesized relationships.
-
References
Cheung, Y.-W. (2007), An empirical model of daily highs and lows. Int. J. Fin. Econ., 12: 1-20. https://doi.org/10.1002/ijfe.303
Groette, O. (2024), The Holiday Effect in Stock Markets: Strategies and Seasonal Insights. Quantified Strategies. Visited 21/11-2024. Link available: https://www.quantifiedstrategies.com/holiday-effect-trading-strategy/
AI Declaration
During the assignment, ChatGPT (Model o1-preview) was solely used to help with writing code in Stata. Specifically for the purposes of generating code to tackle certain operations (e.g. Oil prices futures as an ordinal variable), analyzing error messages when coding to swiftly overcome hurdles, and to fine-tune the graphs used.
Comments
Post a Comment