This study applies four machine learning algorithms (Logistic Regression, Random Forest, Support Vector Classifier, and XGBoost) to predict winners from 1960-2024 US presidential elections with 88.2% accuracy using economic data.
✅ Predicted 2024 correctly (Trump victory)
✅ 15 of 17 elections classified correctly (1960-2024)
✅ Three independent ML algorithms converged to identical 88.2% accuracy
- Gross capital formation (~24% importance)
- Unemployment (14-29% importance)
- Inflation (8-18% importance)
2016 Trump and 1976 Carter were consistently misclassified, which may reflect their unique political circumstances: the rise of populism with Trump and post-Watergate sentiment for Carter.
This analysis explores whether macro socioeconomic indicators can predict US presidential election outcomes. Four machine learning classification models (Logistic Regression, Random Forest, Support Vector Classifier, and XGBoost) were systematically optimised to predict whether the winning candidate would be a Democrat or Republican, based on annual economic and social data during election years.
Sources:
- Election outcomes including winner and party (1960–2024)
- World Bank macroeconomic indicators: GDP growth, inflation, foreign investment, etc.
- Federal Reserve Bank unemployment rates, with monthly data aggregated to yearly averages
Preprocessing:
- Columns with >50% missing values were dropped
- Linear interpolation filled remaining gaps
- Final dataset: 35 features, 17 samples (elections from 1960–2020)
The small sample size (n=17) required Leave-One-Out Cross Validation (LOOCV).
The analysis followed a systematic optimisation process:
- Base: All 35 features with LOOCV
- Selected: 10 features identified by SelectKBest (ANOVA F-statistic)
- Interact: 15 features including interaction terms (products of features)
- Tuned: Hyperparameter optimisation via grid search with LOOCV
Four classification algorithms were evaluated using Leave-One-Out Cross-Validation (LOOCV). Random Forest demonstrated the highest baseline accuracy (70.6%), whereas Logistic Regression achieved superior class separation (ROC-AUC 0.833).
Feature selection:
SelectKBest with ANOVA F-statistic identified the 10 most predictive features, improving performance across all models.
Interaction terms: Products of features (degree=2, interaction_only=True) were trialed, producing 55 generated features from the 10 base features. SelectKBest then reduced these to 15 most predictive terms, capturing multiplicative relationships between indicators. The final models use eight unique base features (four standalone and seven appearing in interaction terms):
- Energy use (kg of oil equivalent per capita)
- Fertility rate, total (births per woman)
- Gross capital formation (% of GDP)
- Inflation, GDP deflator (annual %)
- Immunization, measles (% of children ages 12-23 months)
- Population growth (annual %)
- Revenue, excluding grants (% of GDP)
- Unemployment (%)
Key interaction terms include:
- Population growth × Fertility rate
- Inflation GDP deflator × Gross capital formation
- Measles immunisation × Energy use
- Population growth × Unemployment
A comprehensive grid search with LOOCV was conducted on the four algorithms:
Logistic Regression (408 combinations):
- Best parameters: C=0.001, penalty='l2', solver='liblinear'
- Best CV accuracy: 0.882
Random Forest (216 combinations):
- Best parameters: n_estimators=100, max_depth=3, max_features='sqrt'
- Best CV accuracy: 0.882
Support Vector Classifier (3,060 combinations):
- Best parameters: C=1, kernel='linear', gamma='scale'
- Best CV accuracy: 0.882
XGBoost (22,032 combinations):
- Best parameters: n_estimators=100, max_depth=3, learning_rate=0.05, subsample=0.6
- Best CV accuracy: 0.765
The chart above reveals that hyperparameter tuning contributed minimally to final performance, with feature engineering driving the majority of improvements:
Performance progression by stage:
| Model | Base → Selected | Selected → Interact | Interact → Tuned | Tuning Gain |
|---|---|---|---|---|
| Logistic Regression | 59% → 65% | 65% → 82% | 82% → 88% | 6% |
| Random Forest | 53% → 65% | 65% → 88% | 88% → 88% | 0% |
| SVC | 59% → 65% | 65% → 88% | 88% → 88% | 0% |
| XGBoost | 59% → 59% | 59% → 65% | 65% → 82% | 17% |
Observations:
-
Interaction terms drove performance: The jump from Selected (10 features) to Interact (15 features with interactions) provided the largest gains (+23% for RF/SVC, +17% for LR)
-
Minimal hyperparameter sensitivity: Random Forest and SVC achieved 88.2% accuracy with default hyperparameters; grid search found no improvements, suggesting these models were insensitive to hyperparameter choices on this dataset
-
Overfitting concerns: Extensive hyperparameter searches (216-22,032 combinations) typically yielded minimal or no improvements, suggesting the models (except XGBoost) are not overfitting to hyperparameter selection
-
XGBoost required tuning but underperformed: The more complex XGBoost architecture benefited most from hyperparameter optimisation but still underperformed compared with simpler models, suggesting overfitting on limited data rather than capturing genuine signal
-
Model convergence: Three independent algorithms (LR, RF, SVC) converged on identical 88.2% accuracy after optimisation, suggesting 88.2% represents the performance ceiling achievable with macro socioeconomic indicators alone
This analysis demonstrates that model simplicity combined with feature engineering outperformed complex architectures, strengthening confidence in the 88.2% result as genuine rather than an artifact of overfitting.
The table below shows model predictions with probability of Republican victory in brackets. Probabilities >0.5 predict Republican; <0.5 predict Democratic.
Note on SVC probabilities: SVC uses Platt scaling to calibrate probabilities, which can be unreliable with small datasets (n=17). In some cases, the SVC predicted probabilities contradict the actual predictions (e.g., 1972: predicted Democratic but shows 62.9% Republican probability). This is a known limitation of SVC probability calibration. The tick marks (✅/❌) are based on the actual SVC predictions from the decision function, which are reliable, not the probability threshold.
| Year | Winner | Party | LR | RF | SVC | XGB |
|---|---|---|---|---|---|---|
| 2024 | Donald Trump | Republican | ✅ (0.510) | ✅ (0.675) | ✅ (0.982) | ✅ (0.656) |
| 2020 | Joseph Biden | Democratic | ✅ (0.493) | ✅ (0.180) | ✅ (0.117) | ✅ (0.291) |
| 2016 | Donald Trump | Republican | ❌ (0.498) | ❌ (0.140) | ✅ (0.494) | ❌ (0.276) |
| 2012 | Barack Obama | Democratic | ✅ (0.493) | ✅ (0.282) | ✅ (0.111) | ✅ (0.367) |
| 2008 | Barack Obama | Democratic | ✅ (0.498) | ✅ (0.387) | ✅ (0.405) | ✅ (0.492) |
| 2004 | George W. Bush | Republican | ✅ (0.504) | ✅ (0.705) | ✅ (0.655) | ✅ (0.792) |
| 2000 | George W. Bush | Republican | ✅ (0.510) | ✅ (0.660) | ✅ (0.891) | ✅ (0.557) |
| 1996 | William Clinton | Democratic | ✅ (0.499) | ✅ (0.489) | ✅ (0.529) | ❌ (0.592) |
| 1992 | William Clinton | Democratic | ✅ (0.484) | ✅ (0.146) | ✅ (0.007) | ✅ (0.333) |
| 1988 | George Bush | Republican | ✅ (0.509) | ✅ (0.957) | ✅ (0.790) | ✅ (0.837) |
| 1984 | Ronald Reagan | Republican | ✅ (0.511) | ✅ (0.655) | ✅ (0.771) | ❌ (0.469) |
| 1980 | Ronald Reagan | Republican | ✅ (0.519) | ✅ (0.655) | ✅ (0.811) | ✅ (0.538) |
| 1976 | Jimmy Carter | Democratic | ❌ (0.504) | ❌ (0.880) | ❌ (0.669) | ❌ (0.700) |
| 1972 | Richard Nixon | Republican | ✅ (0.503) | ✅ (0.680) | ❌ (0.629) | ✅ (0.543) |
| 1968 | Richard Nixon | Republican | ✅ (0.505) | ✅ (0.670) | ✅ (0.943) | ✅ (0.674) |
| 1964 | Lyndon B. Johnson | Democratic | ✅ (0.488) | ✅ (0.130) | ✅ (0.284) | ✅ (0.257) |
| 1960 | John F. Kennedy | Democratic | ✅ (0.483) | ✅ (0.120) | ✅ (0.133) | ✅ (0.253) |
1. Model confidence varies dramatically
Logistic Regression shows remarkably low confidence, with probabilities clustering around 0.5 (range: 0.483–0.519) despite 88.2% accuracy. Random Forest and SVC show strong confidence, with probabilities often exceeding 0.65 or below 0.35.
2. The 2016 anomaly
- LR: 0.498 probability Republican
- RF: 0.140 probability Republican
- SVC: 0.494 probability Republican
- XGBoost: 0.276 probability Republican
SVC did correctly predict Republican. However, this is a case where the result from the model's decision function is in tension with its predicted probability (0.494 < 0.50: Democratic), raising questions over the reliability of SVC in this case.
3. The 1976 anomaly
- LR: 0.504 probability Republican
- RF: 0.880 probability Republican
- SVC: 0.669 probability Republican
- XGBoost: 0.700 probability Republican
All models predicted a Republican victory with high relative confidence, which suggests that this election was driven by uniquely political circumstances in the post-Watergate era.
3. Consensus predictions
Elections with unanimous model agreement and high confidence:
- 1988 Bush (R)
- 1992 Clinton (D)
- 2024 Trump (R)
- 1960 Kennedy (D)
4. XGBoost underperformance
Whilst LR, RF, and SVC each made 2–3 errors, XGBoost misclassified 4 elections (2016 Trump, 1996 Clinton, 1984 Reagan, 1976 Carter). The model's more complex architecture may be less suitable to the limited training data, despite regularisation.
5. Ambiguous signals (near-50% probabilities)
- 1996 Clinton: LR=0.499, RF=0.489, SVC=0.529, XGB=0.592 (Clinton's re-election during political scandal)
- 2008 Obama: LR=0.498, RF=0.387, SVC=0.405, XGB=0.492 (Obama's first election, following the 2008 financial crisis)
These near-equiprobable predictions across models suggest the macro socioeconomic indicators provided genuinely ambiguous signals.
Feature importance aggregated across models (interaction terms split equally between base features):
| Feature | LR | RF | SVC | XGB |
|---|---|---|---|---|
| Energy use [kg oil per capita] | 0.060795 | 0.065594 | 0.061296 | 0.065413 |
| Fertility rate [births per woman] | 0.069496 | 0.063800 | 0.058606 | 0.058809 |
| Gross capital formation [% of GDP] | 0.225281 | 0.234555 | 0.236828 | 0.224805 |
| Inflation GDP deflator [%] | 0.140208 | 0.167735 | 0.081208 | 0.175989 |
| Measles immunisation [% children 12-23 months] | 0.122953 | 0.107791 | 0.106715 | 0.167578 |
| Population growth [%] | 0.152977 | 0.132104 | 0.140949 | 0.108729 |
| Revenue excluding grants [% of GDP] | 0.064235 | 0.085106 | 0.028848 | 0.066266 |
| Unemployment [%] | 0.164055 | 0.143314 | 0.285551 | 0.132412 |
Observations:
-
Gross capital formation emerges as a dominant predictor across all models, particularly for Random Forest (normalised importance ~0.24). Investment in fixed assets and economic confidence strongly correlate with electoral outcomes.
-
Unemployment shows the second-highest importance (0.14-0.16 across models), with consistent rankings suggesting its robust predictive power.
-
Inflation GDP deflator ranks highly in tree-based models (RF, XGB) but shows more modest importance in linear models (LR, SVC).
-
Population growth and fertility rate contribute moderately across all models, capturing demographic trends.
-
Consistent feature rankings across models suggest these indicators genuinely drive electoral outcomes.
Box plots comparing the eight base features used in the final models between Democratic and Republican victories reveal systematic patterns:
Economic indicators:
- Inflation GDP deflator: Republicans show 72% higher inflation (3.82% vs 2.22% mean), which is the largest relative difference
- Unemployment: Democrats show 25% higher unemployment (6.66% vs 5.30% mean)
Demographic Indicators:
- Fertility rate: Democrats show 16% higher fertility (2.27 vs 1.96 mean)
- Population growth: Democrats show 25% higher growth (1.12% vs 0.90% mean)
- Energy use: Republicans 7% higher (7548 vs 7079 kg oil per capita)
- Gross capital formation: Republicans 6% higher (22.8% vs 21.4% of GDP)
- Revenue excluding grants: Republicans 4% higher (18.0% vs 17.4%)
- Measles immunisation: Republicans 3% higher (91.3% vs 88.3%)
Republican outcomes strongly associate with lower unemployment and higher inflation. Democratic outcomes associate with higher population growth and higher fertility.
Macro socioeconomic indicators proved to be strong but imperfect predictors of US presidential election outcomes, achieving 88.2% accuracy over 17 elections spanning 64 years. Three independent algorithms converged on this accuracy, suggesting it represents the ceiling for what economic data alone can predict. Despite limited training data (n=17), over-fitting concerns are somewhat mitigated by the fact that feature engineering, not hyperparameter tuning, drove performance gains from 65% to 88%, with tuning contributing minimally (0-6%) for the best models. The two consistent misclassifications (2016 Trump, 1976 Carter) both involved exceptional political circumstances. Gross capital formation, unemployment, and inflation emerged as dominant predictors, indicating that economic stability and business confidence strongly influence electoral preferences.
Elections are not independent; political momentum and demographic shifts create dependencies not addressed by LOOCV. Cultural factors, foreign policy, candidate quality, campaign effectiveness, and media dynamics are not captured.
- Incorporate polling data, demographic trends, or sentiment analysis from news/social media
- Expand to state-level predictions (50 states × 17 elections = 850 samples)
- Implement time-series aware validation strategies
- Ensemble economic and non-economic features with regularisation
The workflow from raw data to a set of optimised models is contained in main.ipynb. Repetitive tasks are handled by three internal modules:
process_data.py: Processes, cleans, and merges raw data into data/proc/data.csvtrain.py: Contains a configurable class Train which runs the training and outputs standard results (the model, test predictions, truth values, and probabilities)analyse.py: Performs standard analysis on the results, returning statistics such as the confusion matrix, ROC-AUC, and accuracy
The structure of the project is given below:
├── code
│ ├── analyse.py # Model performance analysis
│ ├── main.ipynb # Main project notebook
│ ├── plot.mplstyle # Plot styling
│ ├── process_data.py # Data processing module
│ └── train.py # Model training
├── data
│ ├── proc # Processed data (generated)
│ └── raw # Raw data
│ ├── election_data.csv
│ ├── FRED_unemployment_rate_1948-2024.csv
│ └── world_bank_indicators_1960-2023.csv
├── images # Images (generated)
├── models # Models (generated)
├── README.md
├── results # Analysis results (generated)
Libraries: pandas, numpy, matplotlib, scikit-learn, XGBoost