What's next for Bangladesh's economy
This page publishes forecasts for four key measures of Bangladesh's economy: consumer-price inflation, GDP growth, workers' remittances and electricity demand. We compare a wide range of forecasting methods (statistical, econometric, machine-learning and combinations of them) and publish the one that holds up best, with an honest range around it.
Forecasts last updated . Official sources are checked every day.
We show how uncertain each forecast is and keep score against what actually happens. Forecasts are selected using out-of-sample validation and monitored for bias and stability.
Our forecasts, on the record
One current forecast for each series. We score every one against the first official release, and we never edit a forecast after publishing it.
| Series | Published | Forecast | 80% range | Actual (first release) | Error | Status |
|---|---|---|---|---|---|---|
| Inflation · Jan 2027 | 12 Nov 2026 | 8.27% | 6.80–10.18% | — | — | Waiting for the official number |
| Economic growth · FY27 Q1 (Jul–Sep 2026) | 16 Nov 2026 | 3.7% | 1.52–5.59% | — | — | Waiting for the official number |
| Money sent home · Jan 2027 | 12 Nov 2026 | $2.97bn | $2.55–3.48bn | — | — | Waiting for the official number |
| Electricity demand · Nov 2026 | 20 Oct 2026 | 8.12 TWh | 7.64–8.97 TWh | — | — | Waiting for the official number |
What changed
A dated note each time the forecasts are updated, including what moved and by how much.
- New official data: Economic growth for FY26 Q4 (Apr–Jun 2026)
- Inflation: no change.
- Economic growth, FY27 Q1 (Jul–Sep 2026): 3.7% (was 4.2%).
- Money sent home: no change.
- Electricity demand: no change.
- New official data: Inflation for Oct 2026; Money sent home for Oct 2026
- Inflation, Jan 2027: 8.27% (was 8.34%).
- Economic growth: no change.
- Money sent home, Jan 2027: $2.97bn.
- Electricity demand: no change.
- New official data: Electricity demand for Aug 2026
- Inflation: no change.
- Economic growth: no change.
- Money sent home: no change.
- Electricity demand, Nov 2026: 8.12 TWh (was 8.08 TWh).
- New official data: Inflation for Sep 2026
- Inflation, Dec 2026: 8.34% (was 8.26%).
- Economic growth: no change.
- Money sent home: no change.
- Electricity demand: no change.
- Forecasts rebuilt after a full comparison of methods
- Economic growth, FY27 Q1 (Jul–Sep 2026): 4.2% (was 4.9%). Growth history for FY24 and FY25 was also aligned with the official series.
- Money sent home, Dec 2026: $2.92bn (was $3.44bn). Forecasts now stop at three months, as far as testing supports.
- Inflation: central forecast unchanged at 8.26%. Ranges widened about twofold after testing showed the earlier ones were too narrow.
- Electricity demand: no change.
How the forecasts are madeMethods, data and scoring
What we publish
One central estimate for each series, with an 80% range around it. The range is where we expect the actual figure to land about four times out of five. It will sometimes land outside, and extreme events can fall well beyond it.
How we choose a forecast
We test many approaches on data they had not seen, stepping forward through history the way a real forecaster would. We publish the approach that performs best and stays stable. If nothing clearly beats a simple benchmark, we publish the simple benchmark. Being complicated is never a reason to choose a method.
Watching for bias
We track whether forecasts run systematically high or low, and we do not quietly adjust them to offset past misses. A method that leans one way does not make the cut just because its average error looks smaller.
Honest ranges
Ranges are built from our own past errors, and widened where history shows they would have been too narrow. We only look as far ahead as our track record supports, which is why remittances stop at three months.
The methods we compare
For every series we run the same head-to-head test across a wide library of methods, grouped here by family.
- Simple benchmarks. The latest value carried forward, the same period a year earlier, the average recent trend, and moving averages over recent months or quarters. Every other method has to beat these.
- Exponential smoothing. Simple, damped-trend and seasonal versions that give more weight to recent observations.
- Theta method. A robust time-series method that blends the long-run trend with short-run movements.
- ARIMA family. Autoregressions and ARIMA models, including seasonal versions.
- Trend and calendar regressions. A trend plus month-of-year effects, with a short-term correction for recent errors. For remittances, also the timing of the Eid festivals, which moves from year to year.
- Seasonal growth projections. Last year's value grown at the recent year-on-year rate.
- Regularised regression. Ridge and elastic-net regressions on recent history, which guard against over-fitting.
- Machine learning. Random forests and gradient boosting on recent history and calendar effects.
- Models with outside indicators. The same machine-learning methods with World Bank commodity-price indicators (oil, energy and food) for inflation, remittances and electricity, and mixed-frequency nowcasts for GDP that use early remittance and electricity signals.
- Forecast combinations. The median, a trimmed average, an average of the recent best performers, and an error-weighted blend of all of the above.
We add methods as the data allow. Not yet tested: multivariate systems (VAR and Bayesian VAR), dynamic factor models, mixed-frequency regressions (MIDAS), other gradient-boosting libraries such as XGBoost and LightGBM, and neural networks. They are on the list for when longer histories are available.
What each forecast measures
Inflation
- Measures
- Headline consumer-price inflation: the year-on-year change in the national consumer price index (all items), in %.
- Source
- Bangladesh Bureau of Statistics, published through Bangladesh Bank.
- History
- Monthly from July 2013. Earlier price-index bases are linked to the current 2021-22 base so the series is continuous.
- Looks ahead
- Up to six months.
- In use today
- The latest reading carried forward. No tested method beat it reliably, so we publish the simple benchmark, with a wider range because past ranges were too narrow.
- Inputs tested
- The series' own history, seasonality and World Bank commodity prices.
Economic growth
- Measures
- Real GDP growth: the year-on-year change in constant-price GDP, in %, by fiscal quarter (Q1 is July to September).
- Source
- BBS national accounts, published through Bangladesh Bank.
- History
- Quarterly from FY2016-17 Q1. The latest quarter, FY26 Q4, is the BBS first estimate released on 7 October 2026.
- Looks ahead
- One quarter. This is a nowcast: an estimate made before the official figure is out.
- In use today
- The average of the latest four quarters of growth. Models using early remittance and electricity signals did not beat it.
- Inputs tested
- The series' own history, and remittance and electricity growth as early signals.
Money sent home
- Measures
- Workers' remittance inflows, in US$ million per month.
- Source
- Bangladesh Bank (Monthly Economic Trends). The latest month can be preliminary until Bangladesh Bank publishes it.
- History
- Monthly from July 2014.
- Looks ahead
- Up to three months, as far as our track record supports.
- In use today
- A weighted blend of several statistical methods, which edged out the best simple benchmark.
- Inputs tested
- The series' own history, seasonality, Eid timing, and World Bank oil and energy prices.
Electricity demand
- Measures
- Monthly electricity demand, in terawatt-hours (TWh).
- Source
- Ember monthly electricity data, cross-checked against BPDB generation data where available.
- History
- Monthly from January 2019.
- Looks ahead
- Up to six months.
- In use today
- A trend-and-seasonal regression with a short-term correction. It beat the best simple benchmark by a wide margin at every horizon.
- Inputs tested
- The series' own history, seasonality and World Bank energy prices.
Reading the dates
Monthly figures (inflation, remittances, electricity) are labelled by the calendar month they describe. Growth is quarterly and follows Bangladesh's fiscal year, which starts in July: Q1 is July to September, Q2 October to December, Q3 January to March and Q4 April to June. So FY27 Q1 means July to September 2026. "Looking ahead" counts from the latest official figure, so a month that has ended but has not been released yet is still a forecast. Each card lists the full path, month by month; the ringed point and the large number are the headline forecast. For growth, the quarter that has just ended is a nowcast: an estimate made before the official figure is out.
How a forecast is tested
- Out of sample, step by step. We go back through history in expanding windows (rolling origin). At each step a method sees only the data that existed at that point, forecasts the following months or quarters, and is scored against what then happened. Calendar effects such as Eid are known in advance, so they are allowed. Nothing from the future leaks in.
- Error and loss. Error is forecast minus actual, so a positive error means we were too high. Methods are ranked on mean absolute error, which is not dominated by a few extreme misses. Squared-error measures are reported but do not decide the ranking.
- Enough tests. A method needs at least 24 test points for a monthly series and 14 for GDP, and cannot fail to produce a forecast more than 10% of the time. In practice we have around two dozen tests for GDP, 50 to 90 for electricity and remittances, and about 100 for inflation.
- Each horizon separately. One month ahead and six months ahead are different problems, so every horizon is tested and chosen on its own.
How the published method is chosen
The selection rule was written down and frozen before we looked at the results, so the winner is not fitted to the scoreboard.
- Bias screen. A method is dropped if its average error is large relative to its typical error (above 0.40 of it), or if it misses in the same direction more than 70% of the time.
- Simple unless proven otherwise. If the best method is within 3% of a simpler one, we take the simpler one. A more complex method must beat the best simple benchmark in a one-sided Diebold–Mariano test (Harvey–Leybourne–Newbold small-sample version) at the 10% level, and must not be worse than the benchmark by more than 5% in either half of the test period.
- Combinations next. If no single method is clearly better, we publish the best combination of methods when it is at least as accurate as the benchmark and stable across both halves.
- Otherwise the benchmark. If none of that holds, the best simple benchmark is the forecast.
- Disagreement is information. If the best methods disagree widely about today's forecast (a spread of more than twice the selected method's typical error), we flag the forecast as fragile instead of averaging the disagreement away.
After seeing the first results we made three refinements, which we record here because they came after the fact: prefer the same method across horizons when accuracy differs by under 2%; keep one method across all horizons when that gives a coherent path (used for inflation); and calibrate ranges across horizons together rather than one at a time. We will re-validate these choices once each series has about six scored forecasts.
How the ranges are built
The 80% range is the forecast plus the 10th and 90th percentiles of that method's own earlier errors at the same horizon, with at least 12 errors needed. We do not assume errors follow a bell curve, so a range can be lopsided. Where testing showed such ranges were too narrow, we widen them by a common factor, capped at 2. Inflation ranges are widened to twice their raw width, which brought test coverage to roughly 72% to 83% across horizons. A horizon is published only if coverage in testing falls between 70% and 90% and the method is no worse than the benchmark. That is why remittances stop at three months and growth at one quarter.
The limits of forecasting
We take seriously the critique of economic forecasting that single-number predictions overstate what is knowable, especially for variables with rare, very large moves. We do not treat any one view as the final authority on which method to use. We turn the useful lessons into rules.
- A point forecast is not knowledge. We always publish a range with the number, and we avoid false precision.
- Errors are not bell-shaped. Growth and remittance changes have fat tails: rare, very large misses happen. Ranges come from actual past errors, not from a normal-distribution formula.
- A range is not a worst case. The 80% range will miss about one time in five, and a real shock can land far outside it. We track the largest past miss separately as a stress reference and never present it as a probability.
- Robustness over a fragile gain. A method that is slightly more accurate on average but leans one way, or only works in half the history, does not win.
- No automatic bias adjustment. We report bias but never quietly shift a forecast to offset past misses. An adjustment would have to pass its own out-of-sample test, use only earlier errors, be shrunk towards zero and be capped.
- Shorten the horizon when evidence weakens. We would rather say less than claim what the record cannot support.
- Archive every miss. Every forecast stays on the record and is scored.
What the tests showed
- Electricity: trend-and-seasonal models beat the best simple benchmark by roughly a third at every horizon, a clear and statistically solid gain.
- Inflation: no method reliably beat carrying the latest reading forward, so that is what we publish, with ranges widened because the raw ones were too narrow.
- Remittances: no single method significantly beat a three-month average. A blend of methods was slightly better. Projecting last year's seasonal pattern forward was clearly worse.
- Growth: nothing beat the average of the latest four quarters, including models using early remittance and electricity signals. Growth has had occasional very large misses, so we treat it as a nowcast and keep the range wide.
- Outside indicators: World Bank oil, energy and food prices added no accuracy out of sample. Longer yearly World Bank histories (decades of inflation, growth and remittances) pointed the same way: simple methods were hard to beat.
Honest limits: histories are short (electricity starts in 2019, quarterly growth in 2016-17), tests use the latest revised data rather than what was first published, and the live record is only beginning. Scoring against first official releases is the real test, and it is what this page will show over time.
Data and scoring
We use official releases: the Bangladesh Bureau of Statistics and Bangladesh Bank for prices, growth and remittances, and Ember for electricity. Older price-index bases are linked to the current base so inflation forms one continuous series. When BBS has published an annual growth figure but not yet the matching quarter, that quarter is implied from the annual figure and marked with an asterisk until the quarterly release arrives. The latest month of remittances can be preliminary. World Bank commodity-price and annual data are used for testing outside indicators. We score each forecast against the first official number. Later revisions are kept separately and never rewrite the score.
Updates and the record
Sources are checked every day, and the forecasts change only when new official data arrive. Automatic checks flag gaps, implausible jumps and large revisions, and every update is reviewed by hand before it goes live. Each version is archived with its date and never edited, and every miss stays on the record.