Financial Stress Bulletin — 2026-10-08

• Model applied: N3 — volatility-filtered historical simulation + bridge equations for delayed data. Selected among 4 candidates (N1–N4) by error over the 01/2009–12/2018 validation period. The uncertainty range comes from N4 — N3 + mean reversion for VIX, MOVE and credit spreads (lowest interval score on the validation period).
• Tests (test period 01/2019→10/2026, 1940 days; rules registered 07/10/2026)
  • Vietnam: mean absolute error 1.16 points vs 1.37 points when keeping the observed value (15% lower); Diebold–Mariano p < 0.001 → passed.
  • Vietnam: the 80% range covered 87% of days (target 80%); Christoffersen conditional test p < 0.001 → failed.
  • Global: mean absolute error 1.58 points vs 2.10 points when keeping the observed value (25% lower); Diebold–Mariano p < 0.001 → passed.
  • Global: the 80% range covered 85% of days (target 80%); Christoffersen conditional test p < 0.001 → failed.
• Optimisation: the model was selected on the validation period; parameters (λ, window, shrinkage) are pre-registered conventions and were NOT tuned on test results; recalibrated quarterly (latest 07/10/2026).
• The error log has just started; not enough data yet (needs ≥ 5 days with outcomes).

How: (1) series published with a delay are brought up to the end of yesterday by regressions on same-day market series; (2) today's moves are drawn at random from the last 1250 days of history, scaled by current volatility (EWMA, λ = 0.94); (3) each of 2000 scenarios is passed through the exact current index procedure; (4) the point estimate is the median, ranges are the 10%–90% (80%) and 2.5%–97.5% (95%) percentiles.

Series bridged today: Brent crude (2 working days), BAA–10Y spread (2 working days), broad USD index (4 working days), VIX (2 working days).

• Stress episodes (VIX > 30, 367 days): error 17% lower (Vietnam).
• Stress episodes (VIX > 30, 367 days): error 20% lower (Global).

Candidate models (mean absolute error, points; validation → test)

Limits of the estimate

• The estimate uses market data through the end of yesterday only; it does not use the macro calendar, news or scenarios — unexpected events during the day cannot be predicted (they enter only indirectly via current volatility in the range width).
• Weekly/monthly/quarterly/annual series (EPU, GPR, CPI…) keep their latest values; some domestic series may be updated late.
• Tests use revised data and source delays measured on 7 Oct 2026; the lag between the US and Vietnamese sessions within the same day is not modelled.

The 10/2026 point on the chart is the index computed so far within the month (provisional) and only becomes official when the month ends (31/10/2026). The estimate below is the expected 10/2026 index: it is recomputed every working day from new data and converges to the month's index as data completes (17 working days left including today; the range narrows over time).

• Vietnam: the figure is indicative only in the 16–23-working-days-left group — the model has not significantly beaten the "keep the observed value" benchmark in both the validation and test periods; read it together with the range (dashed diamond on the chart).
• Global: the figure is indicative only in the 16–23-working-days-left group — the model has not significantly beaten the "keep the observed value" benchmark in both the validation and test periods; read it together with the range (dashed diamond on the chart).
• Vietnam: the 80% range did not pass calibration for the 16–23-days-left group (test period: actual coverage 63% of days; month-level Christoffersen test p < 0.001); treat it as an indicative range, not an exact probability.

• Model: N3 — volatility-filtered historical simulation + bridge equations for delayed data. The uncertainty range comes from N3. Selected among 4 candidates by error over the 01/2009–12/2018 validation period.
• How: simulate the path for each working day from today to the last working day of the month (each day's shock is drawn at random from the last 1250 days of history, scaled by current volatility; delayed series are bridged by regression), then pass each of 2000 paths through the exact index procedure. Monthly, quarterly and annual indicators keep their current values; the month's own Google Trends is not yet known, so the prior month's value is carried.
• Tests for the 16–23-working-days-left group (test period 01/2019→09/2026, 93 months; month-level tests, rules registered 07/10/2026)
  • Vietnam: mean absolute error 6.77 points vs 7.05 points when keeping the observed value (4% lower); month-level Diebold–Mariano p = 0.170 → failed.
  • Vietnam: the 80% range covered 63% of days (target 80%); month-level Christoffersen p < 0.001 → failed.
  • Global: mean absolute error 6.66 points vs 7.01 points when keeping the observed value (5% lower); month-level Diebold–Mariano p = 0.087 → failed.
  • Global: the 80% range covered 75% of days (target 80%); month-level Christoffersen p = 0.110 → passed.
• Optimisation: the model was selected on the validation period; parameters (λ, window, shrinkage) are pre-registered conventions and were NOT tuned on test results; recalibrated quarterly (latest 07/10/2026).
• The monthly-estimate error log has just started; outcomes are only known after the month closes.

Limits of the monthly estimate

• The estimate targets the month index using the information available when the month ends and does NOT include later revisions — the official monthly index may be revised: when monthly indicators are published after the month ends (notably the month's own Google Trends), the Vietnam point is revised by about 4.5 points on average (90th percentile about 11) and the global point by about 0.9 — past monthly points on the chart are revised figures.
• Early in the month the range is very wide and narrows as the month's data accumulate; in many groups the model has not beaten the "keep the observed value" benchmark, so the point figure is indicative only and must be read with the range. The Vietnam index is also affected by monthly indicators that are not simulated, so its range often fails calibration.
• The limits of today's estimate (no news or events calendar, revised data, source delays measured on 7 Oct 2026) also apply.