How happy is the world, and where is it heading?
Every March the World Happiness Report asks people in about 150 countries to rate their lives from 0 to 10. Here you can explore those answers, compare countries and see our forecasts through 2030.
The forecast
Where each country stands, where we think it is heading, and how far to trust that.
Latest World Bank data
Fetches the latest GDP from the World Bank and updates the estimate for this year with a small ridge model that runs in your browser. If the World Bank cannot be reached, the estimate stays on the published data and says so.
3D globe World Bank update
Drag to rotate in 3D • Click any nation to predict instantaneous well-being shift
Quick nowcast
WHR empirical benchmark vs contemporaneous World Bank indicator stream
Tested error:
0.158
Coverage: 155 countries and territories
Dimensions: 6 Core Decomposed Life Dimensions
Latest data and estimates (155 Territories)
Live comparative table with latest World Bank GDP per capita, calculated delta, and dynamic ML score
| Rank | Country / Territory | Region | WHR Baseline | World Bank GDP | Reported Year | Quick nowcast | Dynamic Delta (Δ) | Action |
|---|
Our forecast for the 2027 report
We published these numbers on Zenodo on 25 October 2026, before the report comes out, so nobody can adjust them later. When it arrives we will score them here, right or wrong.
Ranges are 95 percent intervals. File fingerprint (SHA 256):
How good are these forecasts?
Testing on countries the model never saw, one report ahead, the typical error is about 0.172 points on a 10 point scale. Most of that skill comes from one simple fact: this year's score looks a lot like last year's.
We expected the neural network to win. It tied with a plain ridge regression. On the 2019 to 2025 data T‑HEIN had an error of 0.172 and ridge had 0.173, while repeating last year's score gave 0.178. On the longer 2008 to 2025 data with yearly survey scores, T‑HEIN reached R² = 0.9874 (error 0.1154), significantly better than every model except ridge (R² = 0.9873). With the yearly scores hidden, T‑HEIN beat ridge in all 15 tests (R² = 0.9737 against 0.9722). Removing the neural parts changed little, and most of the accuracy comes from the linear part of the model.
Beyond one report the picture is humbler. With published scores alone, the direction of change soon drifts towards a coin flip, as the table below shows. With yearly survey scores it stays right for 87 percent of countries two reports ahead, and from three reports ahead repeating the latest score is as accurate as any model. The forecasts registered on this site came from an earlier version of T‑HEIN without yearly scores, and they stay frozen. The results tables are further down this page.
Explore the data
Ask a question, watch the years go by, compare countries, or decide what matters to you.
Ask a question
Questions about the dataAsk any question across 155 countries and territories, six factors, and 2026–2030 T-HEIN projections — instant synthesis at one place
Time Series Animation
Watch happiness evolve over the years from the dataset — continuous chronological playback across 155 countries and territories spanning 2019–2025 observed Gallup survey waves through 2026–2030 T-HEIN forecasts.
Who moved the most?
Averages hide a lot. Pick two years and see which countries rose and which fell.
Compare two countries
Select any two nations for a side-by-side pillar duel, margin tug-of-war, and overlapping radar footprint showdown.
Bilateral Well-being Comparison
Select any two nations for a side-by-side pillar duel and overlapping radar showdown
Regional Aggregates & Trajectory Race
Compare aggregate happiness across ten world regions and watch dynamic leaderboard sprint races.
Regional Happiness Comparison
Average happiness scores across global geographical zones
Country Leaderboard Trajectory Race
Watch top countries race to their well-being ratings
Build your own index
What matters most for societal progress? Assign your personal weights to the 6 core dimensions. Watch all 155 countries dynamically re-rank in real time with live rank delta indicators.
| Rank | Country | Custom Score | Rank Delta | WHR 2025 | Factor Share | Action |
|---|
What-If Policy Scenario Simulator (T-HEIN Prediction Lab)
Perturb national macroeconomic and institutional indicators to observe out-of-sample forecast shifts derived from the T-HEIN model parameters (Thakur & Singh).
Policy Parameter Sandbox & Counterfactual Shifts
Perturb individual national factors and observe causal forecast impacts via live ML inference
Explore by What Matters to You · Dimension Explorer
What do you value most in life? Select a mood below to see countries that excel in that area and explore their stories.
What Does This Mean for You?
Behind every number is a person — someone with dreams, challenges, relationships, and moments of joy. The data shows us patterns, but your own life is uniquely yours. What makes you happy? Is it the people around you? Your health? The freedom to make your own choices?
Connection
Having someone to count on is consistently one of the strongest predictors of wellbeing.
Health
Physical and mental health form the foundation for a good life.
Freedom
The ability to make life choices that matter to you.
Prosperity
Economic security that provides stability and freedom from basic chronic scarcity.
Generosity
Living in a community of kindness where people actively help and support others.
Trust
Living in an honest society with clean governance, high public integrity, and mutual trust.
Factor Distribution Heatmap & Macro Pillars
Matrix comparison across leading economies and well-being pillars across 155 countries and territories.
Factor Distribution Heatmap
Matrix comparison across leading economies and well-being pillars
Interactive Correlation Explorer & Scatter Analyzer Pearson r & OLS Fit
Select any two well-being factors to analyze linear dependency, statistical significance, and structural outliers across 155 countries
Regional Treemap & Continental Cluster Explorer
Nested hierarchical visualization of all 155 countries and territories partitioned into continental clusters. Cell area scales by Log GDP per capita or uniform weights, with chromatic temperature reflecting Cantril ladder well-being.
What we noticed
A few patterns from twenty years of Gallup World Poll data.
Nordic Hegemony: Multi-Year Well-being Stability
Finland, Denmark, and Iceland maintain top-3 stability. Decomposition shows baseline social safety nets (95%+ social support) insulate these nations from macroeconomic shocks.
Latin American Well-Being Outperformance Paradox
Costa Rica (#12), Mexico (#25), and Uruguay (#26) report happiness scores +0.42 to +0.65 points higher than econometric models predict based on GDP per capita alone, driven by dense community bonds.
Steepest Trajectory Contractions (2019–2025)
Afghanistan (1.72) and Lebanon (3.29) display unprecedented multi-year shocks (-1.82 pts), driven by collapse in institutional trust and severe economic contraction.
Central & Eastern European Convergence Fast-Track
Serbia (+1.35 pts), Bulgaria (+1.18 pts), and Romania (+1.05 pts) show the world's strongest sustained upward trajectories over the decade, closing the historical gap with Western Europe.
How we did it
The results in full, the equations, the data, and the people behind this.
What is in our paper
Forecasting National Happiness with a Residual Gated Temporal Expert Network, by Ashish Kumar Thakur and Sunil Kumar Singh. The tables and figures here come from the paper and use the same numbers. The registered hypotheses at the end belong to the Zenodo registration.
Research questions
- RQ1. Does a country's three year history improve prediction for unseen countries and future reports, and do nonlinear experts add anything beyond a linear model?
- RQ2. Do the Gini index, internet use, PM2.5 and ten further indicators add information beyond the six report factors?
- RQ3. How does the model weigh past scores against changes in the factors, and do factor attributions differ between higher and lower income countries?
- RQ4. What does the three year averaging of the published score imply for forecasting, and do yearly survey scores improve forecasts?
Table I. Comparison with earlier studies
| Study | Data | Models | Validation | Persistence benchmark |
|---|---|---|---|---|
| Dixit et al. (2020) | WHR data, 156 countries | GRNN, OLS, other regressors | Train 2016 to 2018, test 2019 | No |
| Xiong (2023) | WHR 2018 and 2019 | KNN, RF, LR, ensemble | 80 to 20 split | No |
| Hamja et al. (2025) | WHR 2018 to 2023 | ML, DL, stacking and blending ensembles | 80 to 20, 70 to 30 and 50 to 50 partitions | No |
| Thakur and Singh (2026) | WHR 2024 cross section | HEIN and 27 baselines | Held out test countries | No |
| This study | WHR 2026, World Bank and Gallup data, 2019 to 2025 and 2008 to 2025 | T‑HEIN and nine baselines | Country grouped 5 × 3 cross validation, rolling origin and out of time backtest | Yes |
Table II. Variables of the main panel
Table III. Further variables in the long panel
Equations
(1) Published score, a three year average of survey waves
$$y_{i,t}=\frac{s_{i,t}+s_{i,t-1}+s_{i,t-2}}{3}$$(3) Change between consecutive scores
$$y_{i,t}-y_{i,t-1}=\frac{s_{i,t}-s_{i,t-3}}{3}$$(4) Softmax gate
$$g_k(x,D)=\frac{\exp(u_k^\top\phi(x,D))}{\sum_{j=1}^{3}\exp(u_j^\top\phi(x,D))}$$(5) T‑HEIN prediction
$$\hat{y}_{i,t}=w^\top x_{i,t}+b+\alpha\sum_{k=1}^{3}g_k(x_{i,t},D_{i,t})f_k(D_{i,t},x_{i,t})$$(6) Training loss
$$\mathcal{L}=\frac{1}{n}\sum_{i=1}^{n}(y_i-\hat{y}_i)^2+\lambda\|\theta\|_2^2$$(7) Window persistence
$$\hat{y}_{i,t}=\frac{2s_{i,t-1}+s_{i,t-2}}{3}$$(10) Nadeau and Bengio corrected t test
$$t=\frac{m_d}{\sqrt{\left(\frac{1}{k}+\frac{n_{\text{test}}}{n_{\text{train}}}\right)s_d^2}}$$Fig. 2. Methodology
Fig. 3. Architecture of T‑HEIN
Table IV. Configuration of T‑HEIN
| Component | Configuration |
|---|---|
| Sequence input | 3 years × 7 features (× 8 on the long panel), each centred on its own mean, scaled on the training fold |
| Summary input | 9 inputs in the forecast, 15 in the nowcast (12 and 18 on the long panel) |
| LSTM expert | Bidirectional LSTM, 16 units per direction, Dense 16 (ELU), Dense 1 |
| Attention expert | Self attention, 2 heads, key dimension 8, average pooling, Dense 16 (ELU), Dense 1 |
| Dense ensemble expert | Three branches on the summary and centred sequence: Dense 16 (ELU), dropout 0.1, Dense 1; outputs averaged |
| Softmax gate | Dense 16 (ELU) on the summary and centred sequence, Dense 3 with softmax |
| Linear backbone | Started at the OLS solution on the inner training rows |
| Regularization | L2 penalty 0.001 on expert and gate weights; residual scale starts at 0.1; expert outputs start at zero; target divided by the OLS residual SD |
| Optimizer | Adam, learning rate 0.001, gradient clip norm 1.0, batch size 16, up to 2,000 epochs |
| Stopping and averaging | Patience 50, learning rate halved after 15 epochs, mean of 3 seeds |
Algorithm 1. Training and evaluation of T‑HEIN
- Build the three year sequences, their centred form and the summary inputs from the panel.
- Split by country into 5 folds, repeated 3 times, so no country is in both training and test.
- Fit standardization on the training countries only, and apply it to both sets.
- Hold out 20 percent of training rows for early stopping.
- For each of three seeds, start the backbone at OLS on the inner training rows, set the expert outputs to zero, train all weights with Adam, and restore the weights with the best validation error.
- Average the seeds and score the result on the test countries.
Metrics
| Metric | Meaning |
|---|---|
| RMSE | Typical error in ladder points; lower is better |
| MAE | Average absolute error; less sensitive to outliers |
| R² | Share of variance explained; higher is better |
| Direction accuracy | Share of changes above 0.2 points with the right sign |
| Interval coverage | Share of scores inside the 95 percent interval |
Table V. Model comparison on the main panel
Table VI. T‑HEIN against each baseline on the main panel
Fig. 4. Error of the models on both panels
Table VII. Model comparison on the long panel with yearly scores
Table VIII. T‑HEIN against each baseline on the long panel
Fig. 5. What the three year average does to the score
Table IX. Yearly scores hidden and shown, on the same rows
Table X. Rolling origin evaluation
Fig. 6. Error by target year
Table XI. Positive and negative controls
Table XII. Removing one part of T‑HEIN at a time
Fig. 7. Random splits flatter a model
Table XIII. Effect of adding further variables
Table XIV. What T‑HEIN learned
Fig. 8. Factor attributions by income group
Table XVI. Accuracy from one to five reports ahead with yearly scores
Fig. 9. Accuracy by horizon
Fig. 10. Outlook for six countries
Table XVII. Robustness of the forecast
Registered hypotheses for the 2027 report
- H1. T‑HEIN has a lower RMSE than persistence.
- H2. T‑HEIN and ridge regression stay within 0.02 RMSE points of each other.
- H3. The 95 percent intervals cover 90 to 98 percent of countries.
- H4. Direction accuracy exceeds 60 percent for countries that move by more than 0.2 points.
These forecasts came from the first version of T‑HEIN, before the training and input changes described in the paper and without yearly scores. They stay frozen as the valid registration.
Conclusions
Comprehensive Country Rankings & Factor Database
Filterable, searchable panel database covering 155 territories with full factor decomposition and 15-fold model validation metrics.
Country Rankings & Indicator Matrix
Complete searchable index of all 155 countries with decomposed factor weights
| Rank | Country | Region | Score | GDP per Capita | Social Support | Life Expectancy | Freedom | Generosity | Trust | Action |
|---|
T-HEIN results: forecast and nowcast
From our study: nine models tested on 15 country grouped folds, with countries kept out of training, using the corrected t test and the Wilcoxon test.
Forecast: Next-Year Forecast (Lagged History Only)
Wins count the folds out of 15 in which T‑HEIN had the lower error. p values come from the Wilcoxon test.
Nowcast: Current-Year Nowcast (Same-Year Factors)
Multi-Horizon Accuracy Across WHR Editions
Extended Feature Impacts (Gini, Internet, PM2.5)
T-HEIN Forecast Mechanics & Theoretical Architecture
Mathematical derivations from the research paper (Thakur & Singh): 3-year rolling average mechanical inertia, L2-anchored residual gating, and honest out-of-sample evaluation protocol.
Theoretical Framework (Journal Paper Formulation)
Theoretical derivation of the 3-year rolling average identity, T-HEIN residual softmax gating, and honest cross-edition forecasting
1. Gallup 3-Year Rolling Average Identity & Mechanical Inertia
Published World Happiness Report scores represent moving averages of survey waves: $$\bar{y}_{i,t} = \frac{1}{3}\left(s_{i,t} + s_{i,t-1} + s_{i,t-2}\right)$$
Mathematical Consequence: Exactly 66.7% of next year's score is mechanically determined by observed past waves (\(s_{t-1}, s_{t-2}\)). Only 33.3% depends on the unobserved new annual survey wave (\(\hat{s}_t\)).
2. T-HEIN: Softmax Gating & Residual Neural Hybridization
Softmax gating weights \(g_k(x)\) and residual hybrid prediction function \(\hat{y}\) anchored on an autoregressive linear backbone:
The linear part carries the level, and the gated experts only correct it. In our tests the correction stayed small.
3. Honest Evaluation Protocol: Forecast and nowcast
Eliminating temporal country data leakage through strict time-boundary partitioning:
Scientific Validity: Unlike standard literature that evaluates via random cross-validation (measuring country identity recall), T-HEIN forecasts future WHR editions strictly using past waves.
4. Updating an estimate with new data
In your browserDynamic adjustment mapping contemporaneous World Bank macroeconomic indicator shocks to instantaneous Cantril ladder shifts:
The 0.34 comes from the published GDP contributions and World Bank GDP. Only GDP is updated live; the other five factors stay at their latest published values.
Try your own data & Batch Inference Engine
Upload custom country survey indicators (CSV) to run real-time T-HEIN forecasting inference.
Custom Survey Importer
Upload custom country survey indicators (CSV) to run real-time T-HEIN forecasting inference
Research Datasets (1960–2025) & Zenodo Distribution
Open-access research artifacts with full provenance and an archive on Zenodo (code and data, forecasts).
Open Data Repository
Open-access research artifacts with full provenance and an archive on Zenodo (code and data, forecasts)
Dataset Preview Explorer
About this project
This started as a paper on predicting happiness from a single year of data. Reviewers asked good questions, and rerunning everything on newer data changed how we saw the problem. So we rebuilt the study around one question: could we have called it before it happened?
Forecasts come from T‑HEIN, a small network that reads three years of each country's history on top of a linear model. The longer outlook, from two to five reports ahead, comes from a ridge regression on past scores, because that is the best tested option at those distances. Scores come from the World Happiness Report 2026, based on the Gallup World Poll. The registered 2027 forecasts come from an earlier version of the model, before we added yearly survey scores and changed how it trains. We keep them frozen so that the 2027 report can test them fairly.
The scores are national averages over three years. They say little about differences inside a country, and this site is not affiliated with the World Happiness Report or Gallup.
Data and code
Code, data and results on Zenodo
World Happiness Report data
Cite this work
Thakur, A. K., and Singh, S. K. (2026). Forecasting national happiness: T‑HEIN, honest evaluation and the World Happiness Forecast tool. Zenodo. https://doi.org/10.5281/zenodo.23831010
Say hello
Ashish Kumar Thakur, PDPM IIITDM Jabalpur, 25pcsw02@iiitdmj.ac.in
Prof. Sunil Kumar Singh, Deen Dayal Upadhyaya Gorakhpur University, sksingh.cs@ddugu.ac.in
Citation Formats & Academic Authors
Cite this open-access research portal and connect with the research team at IIITDM Jabalpur and DDU Gorakhpur University.
Citation Formats
Cite this open-access research portal and methodological framework across publications