Capital in Motion

Choose an economy

Flow

1 · What happened

Quarterly flows, annualised % of GDP. Grey line: the driver model's fitted values. Vertical lines: executive-relevant elections (dashed = early). Shaded: 2026 (partial year). Right of the dotted divider: the forecast with 80% and 95% ranges.

1b · Crises, policy and markets

What else was going on

Capital-flow restrictions

Cumulative net tightening measures since 1950 (higher = more restrictive), inflow- and outflow-related. Bergant, Fernández, Teoh & Uribe (2025), from the IMF AREAER; ends 2023.

FX intervention and policy rate

FX intervention (% of GDP; positive = central bank buys foreign currency), IMF dataset to 2024Q4. Policy rate, end of quarter.

Exchange rate against the US dollar

Local currency per USD, index (first month = 100); up = depreciation. IMF exchange rates.

Government bond yield

2 · Why it happened

Drivers, year by year

Each bar splits the year's average deviation from the country's own 2015–25 average into driver contributions (pp of GDP). The dot is the actual deviation. Global = risk appetite, US rates, dollar and oil. Domestic = growth, inflation, external balance, public finances, commodity terms of trade, real exchange rate and interest-rate gap. These are statistical attributions from a fixed-effects model, not causal effects.

Show all eleven driver categories by year

3 · Outlook

Next four quarters

What if global conditions change?

Shifts the path using this country's estimated sensitivity to each global factor, carried forward by its momentum. Sensitivities come from about 44 quarters of data and are shrunk toward the group estimate, so treat the result as a rough order of magnitude.

VIX change0 pts

US policy rate change0.00 pp

Dollar move (one quarter)0%

Oil price move (one quarter)0%

4 · Risk and buffers

How bad could it get, and can the country absorb it?

Flows at risk for

Flows at risk

Chance a sudden stop starts, next four quarters

Can reserves cover a bad year?

Vulnerability indicators

Early-warning score

5 · Peers

How does it compare?

Peers (default: nearest economies by GDP in the same region and group)

6 · Country brief

One-page summary

A plain-text brief built from the numbers above, ready to paste into a note or email.

7 · How reliable is the forecast?

2024–25 backtest

The model was re-estimated with data to 2023Q4 and asked to forecast 2024Q1–2025Q4, then compared with what happened.

Risk monitor: all economies

Latest flows-at-risk and sudden-stop estimates for the next four quarters, with vulnerability flags, crisis events active in 2025–26 and the latest 3-month currency move against the USD (+ = depreciation). Click a column heading to sort, and a row to open that country's analysis. Probabilities rank risk better than they measure it (rolling AUC 0.74). The ranking is the most reliable part.

Group

Find

How to read this tab. Every model was re-fitted at each forecast date using only data available then. Flows-at-risk and sudden-stop methods were chosen on forecasts made in 2019Q4–2022Q4 and are scored here on untouched forecasts made in 2023Q1–2024Q4. The live tool then re-applies the same choice rule to all years. Full details: report/model_card.md.

Point forecasts: real-time rolling test, 2020–2025

Forecasts made at the end of each year from 2019 to 2024 for the next four quarters. Global factors are frozen at the forecast date, and blend weights are chosen using only earlier years. Lower is better (RMSE, pp of GDP annualised).

Flow

Group

Flows at risk: held-out calibration, 2023–24

Share of outcomes below each predicted percentile; a well-calibrated forecast matches the nominal share. Brackets: 90% bootstrap interval.

Sudden-stop starts: held-out test, 2023–24

AUC: how well probabilities rank economies that did and did not see a sudden stop start (0.5 = chance). Brier: overall probability error, lower is better. "Contagion" adds the share of regional peers already in a sudden stop.

Early-warning thresholds: data-driven vs rule of thumb

Thresholds chosen on 2017–21 to minimise the noise-to-signal ratio (false alarms per episode caught), then tested on 2022–24. Net-flow sudden-stop definition.

What moves the downside tail?

Quantile-regression coefficients for the 5th percentile and the median of next-year flows (standardised by each country's volatility), with 95% country-cluster bootstrap intervals.

Exploratory

Which elections hit inflows hardest? (EMDE)

Election-quarter and post-election effects on gross inflows, split by whether the incumbent lost (V-Dem) and whether the election was early. Specified after the main results; does not count toward the pre-registered test.

Exploratory

Do named crises explain the model's surprises?

Unexplained remainder of the driver model by year, scaled by each country's volatility, in years with and without a named crisis (banking, sovereign default, currency crisis, large IMF financing, inflation above 40%).

Exploratory

Policy levers and flows

Next-quarter flows after policy actions, with country and quarter fixed effects and lagged fundamentals. Policy reacts to flows, so these are associations, not effects of policy.

Exploratory

Do markets react to elections?

Measure

Daily exchange rates around elections

Cumulative depreciation beyond the broad US-dollar move, 20 trading days before to 20 after the vote (FRED daily rates, 12 currencies). Band: 95% bootstrap interval.

Confirmatory−0.14

pp of GDP change in gross inflows in the pre-election window. 95% CI −1.59 to 1.30, Holm p = 1.00.

Confirmatory4 / 4

primary hypotheses not supported: inflows, outflows, rule-of-law moderation, AE vs EMDE difference.

Secondary−1.4

pp: gross inflows and outflows both fall in the two quarters after elections. Net flows do not move. Fragile.

Post-hoc−2.15

pp in EMDE gross inflows in the election quarter itself. Placebo p < 0.001. Needs out-of-sample confirmation.

Gross capital flows by country group, 2015–2025

Aggregate flow ÷ aggregate GDP of reporting economies. Inflows = net incurrence of liabilities (L_NIL_T); outflows = net acquisition of assets (A_NFA_T).

How to read the evidence

Confirmatory Tested exactly as written in the protocol, frozen before any regression was run, with multiple-testing adjustment.

Secondary Pre-registered but outside the primary family. Read as supporting evidence.

Post-hoc Specified after results were seen, to compare with the closest prior IMF study. Treat as a lead to test on new data.

Sample

Economy

Series

Quarterly flows as annualised % of GDP (100 × 4 × flow ÷ annual GDP), not winsorised. Vertical lines mark national elections: solid = scheduled, dashed = early (rule-based), grey = non-executive contest. The shaded band is 2026: partial year, projected GDP denominator, excluded from all estimates.

Elections in the event file, 2015–2026

Country × year

Annual values for all 121 economies in the annual sample, 2015–2025. A dot marks a year with an executive-relevant national election; a ring marks an early election (rule-based). Click a country for its full profile. 2026 is left out because only part of the year is published.

Measure

Group

Sort

Find

Election years against other years, country by country

For each country: average gross inflows in years with an executive-relevant election minus the average in other years. This is descriptive. Each country has only two or three election years, so single-country differences are dominated by noise. The study's regression estimates are in the Model lab.

Model lab

Two-way fixed effects, previous-year macro controls (growth, inflation, current account, fiscal balance, public debt, log GDP per capita), standard errors clustered by country. All 128 combinations were estimated with the study's pipeline.

Outcome

Sample

Election windows

Fixed effects

Same specification across samples

The window term that differs most from zero in the chosen specification, re-estimated in each sample.

Event study around executive-relevant elections

Coefficients for each quarter from 6 before to 6 after the election, relative to 4 quarters before (k = −4, marked). Quarters more than 6 from any election are also in the reference group. 95% intervals, clustered by country.

Outcome

Sample

Elections

Pre-specified robustness tests

Each row re-estimates the baseline with one change. Blue = pre-election window (primary). Orange = post-election window (secondary).

Leave-one-out and placebo calendars

Placebo: each country's election calendar shifted by 8–12 quarters, 500 draws. p = share of placebo estimates at least as large in absolute value.

Multiple-testing adjustment

Research report · v1.0 · 11 October 2026

Capital in Motion: elections, institutional stability and international capital flows, 2015–2026

Summary Question Data Method Results Heterogeneity Dynamics Robustness Post-hoc Limitations Conclusion

Summary

We test whether national elections, institutional quality and global financial conditions change cross-border capital flows in 108 economies (quarterly) and 121 economies (annual) between 2015 and 2025. Flows are IMF balance-of-payments data under the BPM6 manual. Gross inflows are new liabilities to non-residents; gross outflows are new foreign assets acquired by residents. Election dates are triangulated across Wikipedia's national electoral calendars, International IDEA and V-Dem.

Pre-registered result. Gross inflows and outflows do not change in the two-quarter window covering the election quarter and the quarter before it. Institutions, development level and asset class do not change that conclusion. All four primary tests have Holm-adjusted p = 1.00.

Two patterns sit outside the confirmatory tests. After elections, gross inflows and outflows fall together by about 1.4 pp of GDP while net flows stay flat. That pattern is directionally stable but statistically fragile. When the window is narrowed to the election quarter alone, matching the closest prior IMF study, inflows to emerging and developing economies fall by about 2.15 pp of GDP, roughly a third of their average level. That result passes a placebo test and seasonality controls, but it was specified after the main results were known.

Question and hypotheses

The push–pull framework explains capital flows by global factors and domestic fundamentals, and rarely includes dated political events. Firms delay irreversible investment around elections (Julio & Yook, 2012), US multinationals' FDI to foreign affiliates dips before host-country elections (Julio & Yook, 2016), and an IMF working paper finds gross private inflows to 38 emerging markets fall around elections, 1990–2020 (Arakelyan & Evdokimova, 2025). We asked whether this holds in a broader and more recent sample.

H1a / H1b: gross inflows / outflows differ in the pre-election window.
H2: the association varies with rule of law (fixed at 2014).
H3: responses differ across FDI, portfolio equity, portfolio debt and other investment.
H5: exposure to global risk (VIX) varies with rule of law, reserve cover and short-term debt.
H6: the association differs between advanced and emerging/developing economies.

Data

The country universe follows a fixed rule: World Bank-listed, population of at least 1 million in 2014, complete gross flows for at least 9 of 11 years, and a 2014 governance score. Of 196 candidates, 121 qualify annually (33 advanced, 88 emerging/developing) and 108 quarterly (33 and 75). No country was added or removed by hand.

Election dates needed three sources. International IDEA records about half of recent elections with the year only, so Wikipedia's national electoral calendars supply exact dates and each contest must be confirmed by IDEA or V-Dem to enter the analysis. That leaves 282 executive-relevant elections in 2015–2025: 195 scheduled, 71 early by our rule, 16 unclassified. Validation checks pass: the BPM6 financial-account identity holds within 0.5% of GDP in every country-period, and a full rerun from the archived raw data reproduces every table byte for byte.

Gross capital inflows by component, emerging and developing economies

% of group GDP. Source: IMF BOP; IMF WEO.

Method

The baseline regresses a flow outcome on a pre-election window (the election quarter and the one before) and a post-election window (the two quarters after), with country and quarter fixed effects and the previous year's macroeconomic fundamentals. Controls are lagged so the model does not condition on channels through which elections act, such as exchange-rate moves. Standard errors are clustered by country, and Driscoll–Kraay, two-way clustered and wild-bootstrap inference are reported alongside. Causal language is reserved for scheduled elections, whose timing is fixed years in advance.

Main results

The confidence interval for H1a rules out pre-election effects larger than about 1.6 pp of GDP, about 0.12 of a within-country standard deviation of quarterly inflows. Inference agrees across methods: country-clustered SE 0.73, Driscoll–Kraay 0.45, two-way clustered 0.68.

Heterogeneity

None of the eight institutional measures, from WGI rule of law to V-Dem judicial constraints, significantly changes the pre-election association. In emerging and developing economies alone the pre-window estimate is −0.90 (p = 0.08), against +0.77 in advanced economies, but the difference is not significant (p = 0.39). Components do not differ (joint test p = 0.43). A higher VIX lowers inflows more where 2014 reserve cover was higher (−0.50 pp per standard deviation, p = 0.09), but nothing in the global-risk family survives adjustment. Sudden stops do not predict unrest, early elections or a change of chief executive.

Dynamics

Gross inflows around scheduled executive-relevant elections, all economies

Reference quarter k = −4. 95% intervals clustered by country.

No quarter in the run-up is distinguishable from zero. Inflows dip in the first two quarters after scheduled elections and rebound in the third, which looks like a shift in timing more than a lasting change. A stacked design with clean comparison countries gives mostly negative post-election estimates but with standard errors of 3–5 pp, too wide to settle the question.

Robustness

Across 34 pre-specified tests the pre-election coefficient stays between −0.94 and +0.27 and is never significant at 5%. Dropping any single country moves it between −0.47 and +0.19, and placebo calendars give p = 0.84. The post-election coefficient is negative in every test, between −0.86 and −3.00, but p ranges from 0.014 to 0.27. It loses significance without financial centres, without 2020, and in the balanced panel.

Post-hoc: the election quarter in emerging economies

This analysis was specified after the confirmatory results were known and is logged in the protocol change log.

Arakelyan & Evdokimova (2025) define the event as the election quarter itself and remove official-sector liabilities from inflows. Re-estimating with their window:

Our pre-registered window merged the election quarter with the quarter before, which is slightly positive, so the run-up null and this effect are consistent with each other and with the IMF finding of no run-up decline. What does not replicate is the IMF's headline moderation: low political stability does not deepen the election-quarter decline here (interaction +0.08, p = 0.95). The EMDE result survives a Bonferroni correction for the six post-hoc models, 1,000 placebo calendars and country-by-season fixed effects. It should still be confirmed on new elections with the window fixed in advance.

Limitations

Quarterly flows are very noisy (within-country SD of 10–20 pp of GDP annualised), so effects smaller than about 0.1 SD cannot be ruled out.
The executive-relevance and early-election rules are approximations.
Not coded: election closeness, coalition changes, government-formation delays.
Only the revised 2026 vintage of the governance indicators is available.
Not used: UNCTAD greenfield and M&A data, and BIS banking statistics.
Policy events are derived from data rather than hand-coded, so policy results are exploratory.

Conclusion

In 2015–2025, capital flows do not move in the run-up to elections across this broad sample. The evidence points instead to a decline in emerging-economy inflows at the moment of the vote and a short-lived gross retrenchment afterwards. The institutional cushioning reported before does not appear. These are statements about association and about scheduled election timing under stated assumptions. They do not rank countries or political systems.

Full report with references: report/final_report.md. Protocol: report/research_protocol.md (git tag protocol-v1.0).

Glossary

How the country analysis works

Driver model. For each country group (advanced; emerging and developing), quarterly flows in 2015Q1–2025Q4 are regressed on last quarter's flow (momentum), global push factors (VIX, US policy rate, change in the US 10-year yield, dollar and oil price changes), last year's domestic fundamentals from the IMF WEO (growth, inflation, current account, fiscal balance, public debt), the change in commodity terms of trade (IMF CTOT) and the real exchange rate (IMF EER), the real interest-rate gap with the US (World Bank; 49% coverage, gaps set to the country average) and the election-quarter and post-election windows, with country fixed effects. Each country gets its own sensitivity to the global factors, shrunk toward the group estimate (empirical Bayes) because 44 quarters is too few to estimate them alone.

Why. With country fixed effects, each quarter's deviation from the country's own average splits exactly into driver contributions plus an unexplained remainder. These are statistical attributions, not causal effects. The model explains only a small share of quarterly movement, which is typical for gross capital flows.

Outlook. The central forecast blends the driver model, the average of the last four reported quarters and the 2015–25 average. Weights are chosen per group and flow by minimising error in a 2024–25 backtest (estimated to 2023Q4). For gross flows the backtest gave the driver model no weight, so the central path rests on the recent level and the long-run average, plus the election effect where that improved accuracy. Ranges come from backtest errors at each horizon, scaled by the country's volatility; 80% ranges contained 79–84% of 2024–25 outcomes. Scenario sliders apply the model's country sensitivities on top of the central path.

Risk. Flows at risk (gross and net) come from quantile regressions blended with each country's history, with tails recalibrated from past errors. Sudden-stop probabilities give the chance that a new episode starts, under a gross and a net definition. The early-warning score uses thresholds chosen by the signals method. Methods were chosen on forecasts made in 2019–22 and scored on 2023–24 (Track record tab). Every live forecast is stored in a ledger so it can be scored later.

Code: src/forecast/. Model card: report/model_card.md. Ledger: results/forecast_ledger/ledger.csv.

Sources

Definitions

Gross inflows: FDI + portfolio + other investment, net incurrence of liabilities (D_F, P_F, O_F · L_NIL_T).

Gross outflows: the same categories, net acquisition of assets by residents (A_NFA_T), reserves excluded.

Net flows: inflows − outflows; positive means a net inflow.

Private inflows (post-hoc): gross inflows minus central-bank and general-government other-investment liabilities and SDR allocations.

Scaling: quarterly flow × 4 ÷ annual GDP × 100, winsorised at the 1st/99th percentiles in estimation.

Election coding

Wikipedia national electoral calendars give dates. A contest enters only if International IDEA or V-Dem also records it in that year.

Executive-relevant: the presidential contest where V-Dem codes the chief executive as directly elected, otherwise the lower-house contest.

Early: more than 6 months earlier than the country's typical 1990–2014 term implies.

Windows: pre = election quarter and the quarter before; post = the next two quarters.