Application & research

NEPK-FX — Neural Prediction of Exchange Rates

How the NEPK formula predicts structural exchange rate direction. With concrete trading signals based on six-month and twelve-month neural network projections.

By Jacobus van Merksteijn · 28 min read · 28 May 2026 · Edition 2

NEPK-FX: neural prediction of exchange rates based on the NEPK formula

Summary

Key conclusions at a glance

This report presents a quantitative analysis of structural exchange rate prospects for twelve countries based on the NEPK indicator (Net External Productive Core). For each of seven EUR pairs, a neural network has been trained that translates macro fundamentals into exchange rate forecasts at 1, 3, 6, 12 and 24 months.

Taiwan has the highest NEPK in the world in 2025 (8.9%), followed by Korea (5.0%), Norway (4.9%) and China (4.8%).
US (1.0%) and UK (1.1%) have structurally weak productive cores — below the Dutch 4.2%.
The network achieves at 12 months directional accuracy of up to 100% (EURKRW) and 86% (EURJPY) on the holdout test.
Strongest trading signals: KRW and CNY appreciation, USD and JPY weakness.
Concrete put/call strategies have been developed for five currency pairs.

Top recommendations — portfolio allocation

Allocation	Trade	Pair	Conviction
30%	EURKRW PUT 12m	EUR/KRW	Highest ★★★★★
25%	EURCNY PUT 12m	EUR/CNY	Highest ★★★★★
20%	EURJPY PUT 18–24m	EUR/JPY	High ★★★★
15%	EURUSD CALL 12m	EUR/USD	Medium-high ★★★
10%	EURTWD PUT 12m	EUR/TWD	Speculative ★★★

Chapter 1

The NEPK Formula: what does a country structurally earn?

NEPK = Export-VA × α × (1 − τ) × φ

Net External Productive Core — a fundamental alternative to GDP

The formula measures not what a country produces, but what it structurally earns: how much net productive value is generated by a country, after deducting fiscal erosion and foreign profit outflows.

Export-VA

Export value added as % of GDP

OECD TiVA domestic value added — not gross exports. Filters out transit trade.

α (alpha)

Share of productive core in the economy

Industry VA as % of GDP (World Bank). The higher, the larger the productive base.

τ (tau)

Effective tax burden on the core

Total taxes + SSC as % of GDP (OECD Revenue Statistics). High τ = erosion.

φ (phi)

Share in national ownership

1 − foreign equity ownership per country. Measure of profit repatriation.

The Dutch NEPK spiral (2025–2050)

For the Netherlands, the NEPK has fallen from around 7% in the 1980s to 4.2% in 2025. Under unchanged policy the NEPK will decline further to 2.5–3% around 2050.

Year	NEPK	Real income (index)	Real estate (index)	Burden/worker (%)
2025	4.2%	100.0	100.0	30.0%
2030	3.9%	99.3	99.6	34.5%
2035	3.6%	98.4	98.5	39.7%
2040	3.3%	97.4	97.3	45.8%
2045	3.0%	96.1	95.5	53.2%
2050	2.7%	94.8	93.7	62.2%

Chapter 2

NEPK Worldwide 2000–2025

For twelve countries the NEPK has been calculated annually for 2000–2025. Sources: Export-VA via World Bank × OECD TiVA DVA share; α via World Bank industry value added; τ via OECD Revenue Statistics; φ via national equity ownership per exchange.

NEPK 2000-2025 for 12 countries — West on the left, Asia on the right — Figure 1 — NEPK 2000–2025. West (left), Asia (right). Source: NEPK-FX Report, May 2026

"Taiwan has the highest NEPK in the world in 2025 (8.9%)" — driven by TSMC's dominance in the global semiconductor industry.

Taiwan

8.9%

Rise since 2010 driven by semiconductors (TSMC)

Korea

5.0%

From ~3% (2000) → 5% (2025) through industrial diversification

Norway

4.9%

Stable 4–5% from oil exports and state stake in Statoil

China

4.8%

Dutch spiral: peak 8.8% (2006) → 4.8% (2025)

Switzerland

3.5–4%

Stable thanks to pharma and low τ

Japan / UK / US

< 2.5%

Structurally hollowed-out core — below 2.5% since 2000

"US (1.0%) and UK (1.1%) have structurally weak productive cores" — even below the Dutch 4.2%. Fundamentally negative for USD and GBP in the long term.

Chapter 3

Macro buffers per country: NIIP, current account and debt

A low NEPK can be temporarily masked by external buffers. A large net international investment position (NIIP) or a structural trade surplus offers a country exchange rate protection — even when the productive core is weak (see: Japan, NIIP +77%).

NEPK, NIIP, current account and government debt per country (2024-2025) — Figure 2 — NEPK, NIIP, current account and government debt per country (2024–2025). Source: NEPK-FX Report, May 2026

NIIP — net international investment position

Country	NIIP % GDP	Character
Norway	+350%	Oil fund — largest creditor
Taiwan	+200%	Life insurance + trade surplus
Switzerland	+110%	SNB reserves + private wealth
Germany	+82%	Decades of trade surplus
Japan	+77%	Masks low NEPK
Netherlands	+55%	Pension funds + multinationals
Korea	+50%	Young creditor nation
China	+14%	Officially low (capital controls)
Italy	+12%	Net positive since 2020
France	−25%	Structural debtor
UK	−31%	Brexit + services economy
US	−90%	Historic low

Chapter 4

Interest rates and inflation: the carry component of FX

The carry component of exchange rates is determined by interest rate differentials. Real rates are however more important than nominal ones: high rates amid high inflation offer no real return. The model incorporates both dimensions in the feature set.

Figure 3 — Policy rate, 10y yield, inflation and real rate per country (2025). Source: NEPK-FX Report, May 2026

Japan

−2.1%

Real 10y — extremely negative, explains yen weakness

NL / Taiwan

slightly neg.

Inflation above yields — negative real rates

FR / IT / US

1.9–2.3%

Highest real rates — capital-attracting but negative NEPK

Korea

pos.

Moderately positive real rate + strong NEPK — most attractive combination

Norway

pos.

Moderately positive real rate + stable NEPK 4.9%

China

pos.

Moderately positive real rate + NEPK 4.8%

Chapter 5

FX strength score: weighted composite ranking

By combining NEPK, NIIP, current account, debt, real rate and NEPK trend into a weighted score, a structural FX ranking emerges. Green = structurally strong; yellow = neutral; red = structurally weak.

Composite score weights

40%

NEPK (productive core)

20%

NIIP (external wealth)

15%

CA (current account)

10%

Real rate

10%

Government debt

5%

NEPK trend (5y)

FX strength score 2025 per country. Green = strong, yellow = neutral, red = weak — Figure 4 — FX strength score 2025. Green = strong, yellow = neutral, red = weak. Source: NEPK-FX Report, May 2026

Chapter 5 (continued)

Two-axis positioning: NEPK versus NIIP

The two-axis plot visualises the combined position of each country: the NEPK (productive strength, x-axis) versus the NIIP (accumulation of external wealth, y-axis). Bubble size reflects the size of the current account; colour indicates the real rate.

NEPK (x-axis) versus NIIP (y-axis) per country. Bubble size = current account, colour = real rate. — Figure 5 — NEPK (x-axis) versus NIIP (y-axis). Bubble size = current account, colour = real rate. Source: NEPK-FX Report, May 2026

Strategic quadrant conclusion: Korea and Taiwan are in the upper right (high NEPK + strong NIIP) — ideal position for currency appreciation. The US is in the lower left (low NEPK + negative NIIP) — a structurally weak profile for USD.

Chapter 5 (continued)

NEPK direction 2020→2025: who is building up, who is shrinking?

The five-year NEPK change (Δ 2020→2025) is a crucial indicator for momentum: countries that build up their productive core see their currency strengthen; countries that shrink undermine their FX position.

NEPK change 2020 to 2025 per country. Norway, Taiwan and Korea build up; NL and UK shrink slightly. — Figure 6 — Δ NEPK 2020→2025. Norway, Taiwan and Korea build up; NL and UK shrink slightly. Source: NEPK-FX Report, May 2026

Builders

Norway, Taiwan and Korea — positive NEPK momentum, extra tailwind for currency appreciation

Stable

Switzerland and Germany — minor change, neutral contribution to composite score

Shrinkers

NL and UK — declining NEPK momentum, negative trend component in FX score

Chapter 6

Neural Network — Methodology

For each of seven EUR pairs (USD, JPY, GBP, CHF, KRW, CNY, TWD), a separate MLP network (Multi-Layer Perceptron) has been trained that translates macro features into expected log-returns at five horizons: 1, 3, 6, 12 and 24 months. 35 models in total.

Architecture

IN Input layer 19 features — NEPK differentials, rates, FX lags, volatility

H1 Hidden 1 64 neurons, ReLU activation, Dropout 0.2

H2 Hidden 2 32 neurons, ReLU activation, Dropout 0.2

H3 Hidden 3 16 neurons, ReLU activation

OUT Output 1 neuron — log-return over horizon H. Optimisation: Adam (lr=1e-3, wd=1e-4), 400 epochs, MSE loss

Feature set (19 inputs)

Category	Features
NEPK differentials	nepk, expva, α, τ, φ
Interest rates	rate_diff + levels of policy and 10y rate
NEPK levels	country and eurozone
FX momentum	log-returns 1m / 3m / 12m
FX trends	moving average 6m / 12m
FX volatility	rolling standard deviation 6m / 12m
NEPK trend	1-year and 3-year trend component

Training regime: 189 months of data (Aug. 2008 – May 2026). Chronologically split: 80% train (2008–2022), 20% holdout test (2022–2026). 10 models per (pair, horizon) with different random seeds — the average yields the central forecast, the spread yields the 80% confidence interval.

Chapter 6 (continued)

Training progress: convergence per pair and horizon

The training curves show the MSE loss (logarithmic scale) over 400 epochs for all seven EUR pairs and five horizons. All models converge stably — there are no signs of overfitting. Shorter horizons (1m, 3m) typically converge faster than longer ones (12m, 24m).

Training curves per FX pair and horizon: MSE loss over 400 epochs for EURUSD, EURJPY, EURGBP, EURCHF, EURKRW, EURCNY and EURTWD — Figure 9 — Training curves per FX pair and horizon (MSE loss logarithmic over 400 epochs). All 35 models converge stably; no overfitting visible. Source: NEPK-FX Report, May 2026

Chapter 6 (continued)

Holdout test performance: directional accuracy

"The network achieves at 12 months directional accuracy of up to 100% (EURKRW) and 86% (EURJPY) on holdout" — these are out-of-sample results on data the model never saw during training (2022–2026).

Holdout test: predicted vs actual (6-month horizon) for KRW, CNY, JPY, CHF and GBP — Figure 7 — Holdout test: predicted vs. actual return (6m horizon). KRW, CNY, JPY: positive correlation visible. CHF, GBP: political/intervention noise dominant. Source: NEPK-FX Report, May 2026

Results matrix: directional accuracy per pair and horizon

Results matrix NN model: directional accuracy heatmap left, predicted change EUR/X heatmap right — Figure 10 — Results matrix NN model. Left: directional accuracy (%) holdout test per pair × horizon. Right: predicted change EUR/X (%). Green = high, red = low/negative. Source: NEPK-FX Report, May 2026

Pair	Dir-acc 6m	Dir-acc 12m	Dir-acc 24m	Verdict
EURKRW	68%	100%	82%	Most reliable
EURCNY	68%	81%	82%	Strong signal
EURJPY	81%	86%	42%	6/12m strong; 24m weak
EURUSD	65%	75%	33%	12m usable; 24m unreliable
EURTWD	54%	64%	30%	Speculative
EURGBP	46%	56%	52%	Politically driven; avoid
EURCHF	22%	61%	15%	SNB interventions; avoid

Chapter 7

FX Forecasts: projections at 1/3/6/12/24 months

FX fan-chart projections: 5 years of history plus projections 1/3/6/12/24 months with 80% confidence interval — Figure 8 — Fan chart: 5 years of historical rates + projections 1/3/6/12/24m with 80% confidence interval (CI). Source: NEPK-FX Report, May 2026

12 months ahead (May 2027)

Pair	Current	12m projection	Δ%	Dir-acc
EURKRW	1761	1736	−1.4% (KRW stronger)	100%
EURCNY	7.90	7.27	−8.0% (CNY stronger)	81%
EURJPY	185.02	192.70	+4.1% (JPY weaker 12m)	86%
EURUSD	1.164	1.199	+3.0% (USD weaker)	75%
EURTWD	36.65	33.16	−9.5% (TWD stronger)	64%
EURCHF	0.910	0.963	+5.9% (CHF weaker?)	61%
EURGBP	0.863	0.879	+1.8% (GBP weaker)	56%

24 months ahead (May 2028)

Pair	24m projection	Δ%	Dir-acc	Reliability
EURKRW	1656	−6.0%	82%	Strong
EURCNY	7.88	−0.2%	82%	Strong
EURJPY	127.44	−31.1% (JPY snapback)	42%	Medium
EURGBP	0.832	−3.6%	52%	Medium
EURTWD	35.21	−3.9%	30%	Speculative
EURUSD	1.053	−9.5% (USD stronger 24m?)	33%	Unreliable
EURCHF	0.964	+6.0%	15%	Unreliable

Note: 24m forecasts with low dir-acc (CHF 15%, TWD 30%, USD 33%) are unreliable due to regime breaks in 2022–2026.

Chapter 8

Five top opportunities: options strategies

The strategies are based on model forecasts and NEPK fundamentals. Recommended horizon: 6–12 months, where the model is statistically strongest. A CALL = right to buy (wins when the rate rises); a PUT = right to sell (wins when the rate falls). A EURUSD PUT wins when EURUSD falls, i.e. when USD strengthens.

Model: 1761 → 1736 (−1.4% in 12m)

Dir-acc 12m: 100% holdout

NEPK Korea: 5.0% vs Germany 2.2%

KRW structurally undervalued due to historical capital-flight fears. NEPK differential of nearly 3 percentage points relative to the eurozone is consistent and growing.

MainBuy EURKRW PUT strike 1,740, 12 months. Premium ~1.5–2%. Break-even ~1,700.
AlternativeLong KRW / short EUR spot or forward. No premium, directional risk.
Risk-definedBear put spread: PUT 1,750 long, PUT 1,650 short. Max loss = net premium.

Model: 7.90 → 7.27 (−8.0% in 12m)

Dir-acc 12m: 81% holdout

NEPK China: 4.8%, low τ, high φ

Yuan structurally undervalued due to capital controls and strongly positive NEPK. With gradual capital account liberalisation, appreciation would accelerate.

MainBuy EURCNY PUT strike 7.80, 12 months. If it lands at 7.27: profit 5–6% after premium.
AlternativeUSDCNH PUT — comparable signal, somewhat more liquid market.
Risk-definedBear put spread: PUT 7.80 long, PUT 7.20 short. Max loss = net premium.

Model: +4.1% (12m) → −31.1% (24m)

Dir-acc 12m: 86%

Real rate Japan: −2.1% (unsustainable)

Classic "yen snapback": BoJ normalisation is inevitable. A real 10y of −2.1% provides no structural support. Long tenor essential — the 12m model shows temporary weakness before the major rebound.

MainBuy EURJPY PUT strike 180, 18–24 months. Long tenor essential to capture the snapback.
AlternativeUSDJPY PUT — lower implied volatility, cheaper.
Risk-definedCalendar spread: sell short CALL (1–3m), buy long PUT (12–24m).

Model: 1.164 → 1.199 (+3.0% in 12m)

Dir-acc 12m: 75%

US composite score: lowest — NIIP −90%, NEPK 1.0%

The US has the weakest structural position of all countries studied: lowest NEPK, historically negative NIIP, negative current account balance. 75% dir-acc at 12m is usable. Caution with 24m — dir-acc only 33%.

MainBuy EURUSD CALL strike 1.20, 12 months. OTM, premium ~0.8–1.2%. Profit above 1.21.
Risk-definedBull call spread: CALL 1.18 long, CALL 1.25 short. Profit capped at ~+6%.

Model: −9.5% in 12m

Dir-acc 12m: 64%

NEPK Taiwan: 8.9% (highest in the world)

Taiwan has the strongest NEPK in the world, but the central bank actively suppresses TWD. Fundamentally strong case — but volatile due to interventions. Wide CI, options expensive due to high implied volatility. Not recommended for 24m (dir-acc only 30%).

MainBuy EURTWD PUT strike 35, 12 months. Wide CI — not for risk-averse investors.
AlternativeUSDTWD PUT — more liquid market, lower spreads.

Avoid — CHF and GBP: EURCHF too unpredictable due to SNB interventions (dir-acc 15–22%). EURGBP politically driven (dir-acc 42–56%). Both pairs are not included in the portfolio allocation.

Chapter 8 (continued)

Portfolio allocation: diversified options positions

The recommended allocation reflects the hierarchy of model conviction and NEPK fundamentals. Five pairs in total; CHF and GBP are excluded due to insufficient directional accuracy.

Allocation	Trade	Rationale	Conviction
30%	EURKRW PUT 12m	100% dir-acc + NEPK differential 3pp	★★★★★ Highest
25%	EURCNY PUT 12m	81% dir-acc + −8% expected appreciation	★★★★★ Highest
20%	EURJPY PUT 18–24m	BoJ normalisation + real rate −2.1%	★★★★ High
15%	EURUSD CALL 12m	75% dir-acc + weakest composite score	★★★ Medium-high
10%	EURTWD PUT 12m	Strongest NEPK in the world; intervention risk	★★★ Speculative

Chapter 9

Conclusions and model limitations

NEPK provides a coherent framework for structural FX direction.
Neural network outperforms the naïve approach at 6–12m for Asian pairs.
Countries with the weakest NEPK (US, UK) are the most likely FX losers.
Taiwan shows the greatest divergence between fundamentals and market price.
24m forecasts are unreliable for CHF, USD and TWD.

Training data: only 18 years (Aug. 2008 – May 2026).
Annual figures + monthly interpolation miss abrupt moves.
No geopolitical shocks in the model (Taiwan Strait, Ukraine, etc.).
Central bank interventions not explicitly modelled.
OECD TiVA data up to 2020 — extrapolated thereafter.

Recommended improvements

LSTM architecture for time dynamics
Ensemble with classical models (PPP, BEER, FEER)
Risk features: VIX, CDS spreads, term premiums
Rolling-window backtest for regime stability
Daily data for shorter horizons (1m)

Disclaimer

Legal disclaimer and model warning

This analysis is for educational purposes only and does not constitute investment advice, a recommendation or an invitation to buy or sell financial instruments.

FX options are complex financial instruments with significant risks, including the total loss of the premium paid. Central bank interventions, geopolitical shocks and regime breaks can abruptly and entirely invalidate model forecasts.

Past model performance is no guarantee of future results.

Always consult a qualified and regulated financial adviser before acting on the basis of quantitative models. The author and Het Open Vizier accept no liability for trading decisions made on the basis of this report.

Data sources (Chapter 10)

World Bank Open Data — exports, industry VA, manufacturing, tax data

OECD Revenue Statistics — tax-to-GDP incl. social security contributions

OECD TiVA — domestic value added in exports (up to 2020)

Wikipedia / BEA / Bundesbank / Eurostat — NIIP per country

FSS / JPX / TWSE / KOSPI / Deutsche Börse — foreign equity ownership

Perplexity Finance — monthly FX rates 2007–2026

Van Merksteijn (April 2026) — Source report NEPK Netherlands 1970–2050

Het Open Vizier

NEPK-FX — Neural Prediction of Exchange Rates

Key conclusions at a glance

Top recommendations — portfolio allocation

The NEPK Formula: what does a country structurally earn?

The Dutch NEPK spiral (2025–2050)

NEPK Worldwide 2000–2025

Macro buffers per country: NIIP, current account and debt

NIIP — net international investment position

Interest rates and inflation: the carry component of FX

FX strength score: weighted composite ranking

Composite score weights

Two-axis positioning: NEPK versus NIIP

NEPK direction 2020→2025: who is building up, who is shrinking?

Neural Network — Methodology

Architecture

Feature set (19 inputs)

Training progress: convergence per pair and horizon

Holdout test performance: directional accuracy

Results matrix: directional accuracy per pair and horizon

FX Forecasts: projections at 1/3/6/12/24 months

12 months ahead (May 2027)

24 months ahead (May 2028)

Five top opportunities: options strategies

Long KRW — PUT on EURKRW

Short CNY weakness — PUT on EURCNY

Long JPY rebound — PUT on EURJPY

Short USD — CALL on EURUSD

Long TWD — PUT on EURTWD

Portfolio allocation: diversified options positions

Conclusions and model limitations

Recommended improvements

Legal disclaimer and model warning

Data sources (Chapter 10)