## Nowcasting GDP Growth for Kenya — Working Paper No. WP/2026/032 (Content unit: wpiea2026032-source-pdf) ## Source details **Canonical URL:** [Nowcasting GDP Growth for Kenya — Working Paper No. WP/2026/032 (Content unit: wpiea2026032-source-pdf)](https://www.imf.org/-/media/files/publications/wp/2026/english/wpiea2026032-source-pdf.pdf) ## Other formats - [Markdown version](/-/media/files/publications/wp/2026/english/wpiea2026032-source-pdf.pdf.md) - [Structured JSON version](/-/media/files/publications/wp/2026/english/wpiea2026032-source-pdf.pdf.json) --- ### Data: composition, transformations, and selection - Goal: build nowcasting predictions using traditional data routinely monitored by the CBK, markets, and media. - Transformations and treatment: - Variables in Table 1 are transformed in year-on-year (y-o-y) log differences except for PMI and visitor arrivals (left in level); survey-based indicators are included in level. - Predictor selection criteria: - Availability to monetary and economic institutions and markets; economic sector coverage; ability to track Kenya’s main economic sectors; comovement with GDP growth. - Dataset composition: - Quarterly GDP series (target variable). - 11 hard and soft official input data recorded at the monthly frequency. - Hard data: agricultural trade in goods (tea and coffee), services (tourist arrivals), production (vehicles produced, electricity generated), remittances, money supply (M2). - Soft data: purchasing managers’ sentiments (PMI). - Rationale for specific series: - Coffee and tea exports: significant in Kenya’s agricultural production and trade. - Diaspora remittances: Kenya’s largest source of foreign exchange; support household consumption and investment. - Visitor arrivals: capture tourism contribution; key source markets include the USA, UK, India, Tanzania, and Uganda. - Electricity and vehicle production: electricity generation reached approximately 13,000 GWh in 2024; installed capacity grew to 3.82 million kW in 2023; demand peaked at 2,362 MW in July 2025; vehicle production examples include 826 units in November 2024 and 978 units in October 2024; target production to 3,270 units by 2030. - Exclusions: - Consumer Price Index not included because volatility of price indicators in Kenya is substantially higher than that of GDP growth. - Key series metadata and contemporaneous correlations (selected entries from Table 1; transformations and exact values preserved): - Stanbic Bank Kenya PMI — Frequency: Monthly — Start — End Dates: Jan 2014 — Dec 2025 — Lag (days): 5 — Unit: Index — Transformation: Level — Correlation (ρy,gdp): -0.01 — Variance share of Common Component: 0.000 — Source: S&P Global - Diaspora Remittances — Frequency: Monthly — Start — End Dates: Jan 2004 — Nov 2025 — Lag (days): 11 — Unit: NSA, Mil. USD — Transformation: y-o-y %∆ — Correlation (ρy,gdp): -0.02 — Variance share of Common Component: 0.210 — Source: CBK - Mobile Money — Frequency: Monthly — Start — End Dates: Mar 2007 — Nov 2025 — Lag (days): 17 — Unit: NSA, Bil. KES — Transformation: y-o-y %∆ — Correlation (ρy,gdp): 0.02 — Variance share of Common Component: 0.254 — Source: KNBS/CBK - Exports of Goods — Frequency: Monthly — Start — End Dates: Aug 1998 — Nov 2025 — Lag (days): 41 — Unit: NSA, Mil. KES — Transformation: y-o-y %∆ — Correlation (ρy,gdp): 0.19** — Variance share of Common Component: 0.299 — Source: CBK - Imports of Goods — Frequency: Monthly — Start — End Dates: Aug 1998 — Nov 2025 — Lag (days): 41 — Unit: NSA, Mil. KES — Transformation: y-o-y %∆ — Correlation (ρy,gdp): 0.36*** — Variance share of Common Component: 0.618 — Source: CBK - Exports of Coffee — Frequency: Monthly — Start — End Dates: Aug 1998 — Nov 2025 — Lag (days): 41 — Unit: NSA, Tonnes — Transformation: y-o-y %∆ — Correlation (ρy,gdp): -0.02 — Variance share of Common Component: 0.028 — Source: KNBS/CBK - Exports of Tea — Frequency: Monthly — Start — End Dates: Aug 1998 — Nov 2025 — Lag (days): 41 — Unit: NSA, Tonnes — Transformation: y-o-y %∆ — Correlation (ρy,gdp): -0.18** — Variance share of Common Component: 0.104 — Source: KNBS/CBK - Money Supply (M2) — Frequency: Monthly — Start — End Dates: Dec 1995 — Sep 2025 — Lag (days): 45 — Unit: NSA, Mil. KES — Transformation: y-o-y %∆ — Correlation (ρy,gdp): 0.11 — Variance share of Common Component: 0.018 — Source: CBK - Electricity Production — Frequency: Monthly — Start — End Dates: Feb 2006 — Nov 2025 — Lag (days): 48 — Unit: NSA, Mil. KWh — Transformation: y-o-y %∆ — Correlation (ρy,gdp): 0.30*** — Variance share of Common Component: 0.073 — Source: KPLC/KNBS - Visitor Arrivals (Persons) — Frequency: Monthly — Start — End Dates: Jan 2007 — Sep 2025 — Lag (days): 48 — Unit: Number — Transformation: Level — Correlation (ρy,gdp): 0.38*** — Variance share of Common Component: 0.069 — Source: KTB/KNBS - Vehicle Production — Frequency: Monthly — Start — End Dates: Jan 2003 — Nov 2025 — Lag (days): 48 — Unit: Number — Transformation: y-o-y %∆ — Correlation (ρy,gdp): 0.35*** — Variance share of Common Component: 0.055 — Source: KNBS - Real GDP — Frequency: Quarterly — Start — End Dates: 2009 Q1 — 2025 Q3 — Lag (days): 99 — Unit: NSA, Mil. KES — Transformation: y-o-y %∆ — Correlation (ρy,gdp): 1.00 — Variance share of Common Component: 0.368 — Source: KNBS - Notes: CBK—Central Bank of Kenya; KPLC—Kenya Power & Lighting Company; KNBS—Kenya National Bureau of Statistics; KTB—Kenya Tourism Board. ***, **, * indicate statistical significance at 1%, 5%, 10% level. Variance share estimated over January 2010 — September 2021. ### Transaction data and mobile-money indicators - Motivation: timely non-traditional transaction data improve capture of private demand; digital payments data were highly effective during COVID-19. - Novelty: first paper known to use digital payments data to nowcast GDP growth in Kenya. - Mobile payments data (Total agent cash in cash out from CBK): - Publication timing: data published on the 17th of each month for the preceding month. - Reach and expansion (May 2025): - Active mobile money agents: 424,404 - Registered mobile money accounts: 85.62 millions - Number of mobile money transactions: 214.51 millions - Transaction value trends: - Value of mobile money transactions dropped by 21.8% y-o-y to 636.21 billion Kenya Shillings (about 3.7% of 2025 GDP) in February 2025. - Value in May 2025: 713.25 billion Kenya Shillings (about 4.2% of 2025 GDP) — still 1.2% lower than the same month in the previous year. - Interpretation: divergence between transaction counts (edging higher) and transaction value (declining) in H1 2025 suggests a shift toward low-value payments and retreat in large-scale use, reflecting higher living costs and slowing household spending. - Market structure and access: - M-PESA market share: edged down to 91% in 2025 Q1 amid increasing competition. - Financial access: over 85% of Kenyan adults have access to financial services mainly through digital platforms in 2025. - Historical behavior: - Mobile money sector plunged in March–April 2020 and recovered by jumping 84% from its trough in April 2020 to February 2021. ### Sampling, real-time vintages, and correlations - Frequency and lag: all indicators except quarterly GDP are sampled monthly and have a maximum delay of about one and a half months. - Real-time vintages: detailed historical real-time datasets for Kenya do not exist; authors reconstructed pseudo data vintages starting from September 2021 to mimic actual data flow and evaluate the model in a pseudo-real-time setting. - Contemporaneous correlations: - Most contemporaneous correlations (ρy,gdp) in Table 1 are economically and statistically significant. - About half of the contemporaneous correlations exceed 0.2. - PMI and remittances have relatively weak absolute contemporaneous correlations but are timely (reporting lags of 5 and 11 days) and lead GDP growth. - Modeling implication: descriptive analysis supports application of a dynamic factor model to nowcast Kenya’s economic activity. ### Nowcasting framework: dynamic factor model (DFM) and estimation - Modeling choice: - Dynamic factor model as in Giannone et al. (2008); full forecast framework where target and predictors are jointly determined and forecasted. - Model decomposition and state-space representation (preserved in source): - yt = ΛFt + εt - Ft = ΓFt−1 + ut, ut ∼ i.i.d. N(0, Ξ) - εt = Θεt−1 + ηt, ηt ∼ i.i.d. N(0, Σ) - Key assumptions: - Comovements summarized by a small set of unobserved common factors Ft. - Idiosyncratic components orthogonal to each other and to common factors at each lead and lag (diagonal Θ and Σ). - Estimation strategy: - Expectation-maximization (EM) algorithm and Kalman filter; initial estimation via principal components. - Number of factors: models run with k = 1 to 5; authors settle on k = 2 as parsimonious; single-factor model often sufficient in many applications but two-factor chosen to capture pandemic dynamics. - Mixed-frequency aggregation: - Quarterly GDP treated as partially observed at monthly frequency; monthly unobserved GDP growth y^m_t,gdp defined and aggregated to quarterly via y^q_t,gdp = y^m_t,gdp + y^m_{t−1},gdp + y^m_{t−2}, with monthly factor structure y^m_t,gdp = λ_gdp f_t + ε_t,gdp and ε_t,gdp = ρ_gdp ε_{t−1},gdp + η_t,gdp. - Monthly factor series extracted with the Kalman filter and plugged into aggregation to obtain quarterly nowcast. - EM estimation for ragged-edge panels: - Initialization: PCA for F^{(0)} and OLS for Ψ^{(0)}. - E-step: compute expected log-likelihood and extract factors using Kalman filter and smoother. - M-step: re-estimate parameters by maximizing expected log-likelihood; iterate until convergence. - Modified M-step formulas and handling of missing observations preserved (Equations 16–24). - Real-time news decomposition: - Nowcast revision identity: E[y^n_{t,n}|Ω_{v+1}] = E[y^n_{t,n}|Ω_v] + E[y^n_{t,n}|I_{v+1}], with news I_{v+1,j} = y^t_{j,n_j} − E[y^t_{j,n_j}|Ω_v]. - Revision expressed as weighted sum of news: E[y^n_{t,n}|Ω_{v+1}] − E[y^n_{t,n}|Ω_v] = Σ_{j∈J_{v+1}} δ_{v+1,j} (y^t_{j,n_j} − E[y^t_{j,n_j}|Ω_v]). - Kalman smoother output provides news I_{v+1} and expectations to compute weights δ_{v+1,j}. - Practical complication: historical data revisions require iterative parameter reestimation. ### Empirical implementation, model selection, and forecasting design - Model choice and sample: - Two-factor DFM chosen based on empirical findings: single-factor fails to capture COVID-19 recession and recovery; ≥2 factors capture pandemic dynamics better; models with more than two factors virtually indistinguishable. - Estimated from January 2010 to September 2021. - Real-time forecasting exercise: - Simulated forecaster runs model in real time every two weeks beginning September 21, 2021. - Model parameters recursively re-estimated at the start of each quarter using the most recent 12 years of data; parameters fixed within the quarter. - Nowcasts generated solely from variables that would have been released by that time following stylized availability calendar. - Pseudoreal-time period: sequential forecasts from September 2021 to March 2024. - Truereal-time forecast: April 2024 to December 2025 (using real-time vintages for series in Table 1). - Prediction frequency in each quarter: - Six nowcasts of the current-quarter GDP growth (two nowcasts on the 7th and 21st day of each of the three months in the quarter). - Six backcasts of the previous-quarter GDP growth. ### Real-time nowcasting results and variable contributions (2024 Q3 — 2025 Q4) - Nowcast schedule and example timeline: - 12 predictions for each quarter; nowcast covers three months of the reference quarter plus three months following it up to GDP publication date. - First update for 2024 Q3 nowcast: July 7 (coinciding with PMI, M2, and first-quarter GDP growth figures). - July 7–July 21, 2024: - Nowcast fell from 4.06% to 3.57% between updates on July 7 and July 21. - Main negative drivers in July: second-quarter growth rate, PMI, and M2. - Broad-based downward revision except visitor arrivals and tea exports. - Largest downward adjustments: imports, exports, and mobile money; followed by electricity production and remittances. - August–October 2024 pattern: - GDP growth decline begins in July, stabilizes in August, picks up in September and October driven by positive readings in imports, exports, vehicle production, visitor arrivals, and remittances. - Subsequent downward revisions as most macro indicators released, reflecting weak digital money, trade, and vehicle production. - Backcasting for 2024 Q3 begins October 2024; backcasted growth stays around 3.87% through end-December 2024. - Predicted GDP growth for 2024 Q3: 4.65%; actual recorded growth for 2024 Q3: 4.2%. - 2024 Q4 nowcast: - Downward revisions continued until January 2025, driven by disappointing trade and digital money readings, then stabilized through March 2025. - 2024 Q4 GDP data release on May 6, 2025 led to a large upward revision from 3.12% to 4.00%. - 2025 Q2 nowcast: - Gradual upward revision from 3.87% in April to 5.18% in September 2025, approaching the actual growth rate of 5.0%. - Primary drivers of the upward adjustment: imports, followed by vehicle production; digital money provided an initial substantial contribution in April 2025. - 2025 Q3 nowcast: - Final prediction for 2025 Q3: 4.74%; actual published growth by KNBS on December 31, 2025: 4.90%. - 2025 Q4 nowcast: - Forecast edged down from 4.61% on October 7, 2025 to 4.16% on December 21, 2025, primarily due to weak digital money readings, before recovering to 4.50% in January 2026. ### Drivers, timing, and impact assessments - Main contributing sectors to nowcast revisions: - Trade variables (exports and imports) are the most significant contributors; exports and imports exerted largest downward influence in July 2024; imports drove upward revisions in August–September 2024. - Digital money (mobile money) played a crucial role in July and August revisions, with a large negative contribution in August 2024. - Production variables (electricity and vehicle production) are important contributors. - Visitor arrivals contribute positively when recovering. - Timing of useful information: - Digital money figures published 15 to 17 days after month-end. - PMI is the timeliest variable; contributes mainly at the beginning of the nowcast quarter with impact fading over time. - Most useful information arrives in the middle of the nowcasting period (bell-shaped pattern); early and late surprises have less impact. - Four key empirical insights (average absolute weekly impact, real time 2022–2025): - Publication of previous-quarter GDP growth significantly affects nowcast updates for both nowcasts and backcasts. - Most important sectors: trade (exports and imports), production (electricity and vehicle), and consumption (digital money). - Positive association between imports and GDP growth: imports act as catalysts via capital goods and intermediate inputs, supporting productivity, industrialization, and export competitiveness. - Mobile/digital money has substantial contribution: facilitates formal and informal sector activity, boosts financial inclusion, consumption smoothing, risk-sharing, and poverty alleviation. ### Predictive performance, nowcast errors, and annual forecasts - General model performance: - Model tracks actual GDP growth closely over time, capturing the 2022 slowdown and 2023 rebound. - Nowcast accuracy improves as more data arrive closer to GDP publication. - Specific examples: - 2024 Q1 nowcasts (January–March 2024): revised up from 5.06% to 6.40% (six nowcasts). - 2024 Q1 backcasts (April–June 2024): remained around 6.30% until official GDP publication in early July 2024 at 4.94%. - 2024 Q2 nowcasts (April–June 2024): ranged from 5.78% to 4.83%; backcasts updated to values between 3.47% and 3.89% until 2024 Q2 GDP growth of 4.6% published on October 2, 2024. - 2025 Q1 nowcasts: revised downward from 4.16% to 3.12%, driven by negative news in mobile money figures; 2024 Q4 GDP data release partly offset negative effects. - 2025 Q2 and Q3 predictions hovered around 5% and aligned with realized figures. - Nowcast errors: - Defined as Error_t = ˆE[y_t,gdp | Ω_T] − y_t,gdp for 0 ≤ t ≤ T. - Errors shrink toward zero as time approaches official GDP data release (errors computed across 15 quarters from 2021Q4 to 2025Q3). - Annual predictions and benchmarking: - Annual nowcasts constructed by aggregating quarterly forecasts updated bimonthly. - Nowcast for 2024: - Nowcasts decreased from 5.60% to 5.07% in first half of 2024 because of negative trade developments, then rebounded and stabilized to 4.72% in 2025 Q1. - Nowcasts were significantly lower than other institutional predictions in 2024 Q1 but gradually converged toward CBK and WEO forecasts, surpassing them slightly in 2025 Q2 by 2 and 21 basis points, respectively. - Nowcast for 2025: - As of reporting date, 10 data points still missing from required 33. - On December 7, 2025 the model predicted an increase in growth to 4.72% in 2025, up from 4.37% in January 2025. - Nowcasts for 2025 were substantially lower than institutional predictions in Q1 but converged toward WEO projections in Q3 and Q4. - Backtesting annual GDP growth: - 2022 backcasts: dropped from 5.94% to 4.94% in 2022 Q3, then gradually to 4.87% in 2023 Q1, close to realized GDP growth of 4.90%; CBK and WEO forecasts remained higher (~5.6% and 5.37% in 2023 Q1). - 2023 backcasts: rose from 3.44% in 2023 Q1 to 5.72% in 2024 Q1; CBK and WEO forecasts ranged between 5% and 6.1%. - Interpretation: CBK and WEO forecasts display stronger inertia; the DFM nowcasts are more flexible and track actual GDP growth more closely as information accumulates. ### Policy relevance and practical implications - Operational schedule: - Run nowcast of current-quarter growth and backcast of previous-quarter growth on the 7th and 21st day of each month. - Policy uses: - Framework produces timely predictions and interpretable updates to inform stabilization measures and help calibrate development policies. - News decomposition shows which incoming data drive revisions and why, aiding policymakers in identifying sectoral shocks and transmission channels. - Flexibility and extension: - Approach can seamlessly include new predictors as high-frequency and non-traditional indicators grow, enabling application to other LICs with infrequent or annual GDP series. *Source: wpiea2026032-source-pdf - Nowcasting GDP Growth for Kenya — Working Paper No. WP/2026/032 (excerpted content units 2.1–3.2–4.1 and annex highlights).* ### 2.1 Data ### 2.1 Data ### 2.1.1 GDP and economic activity data - Goal: build nowcasting predictions using traditional data routinely monitored by the CBK, markets, and media. - Data transformation: - Variables in Table 1 are transformed in year-on-year (y-o-y) log differences to induce stationarity and address seasonality, except for PMI and visitor arrivals series that are already stationary and therefore are left in level. - Survey-based indicators are included in level because they typically already reflect changes in current macroeconomic conditions. - Predictor selection criteria: availability of data monitored by monetary and economic institutions and markets, economic sector coverage, ability to track activity in Kenya’s main economic sectors, and comovement with GDP growth. - Dataset composition: - Quarterly GDP series (target variable). - 11 hard and soft official input data recorded at the monthly frequency. - Hard data: real sector official agricultural variables on trade in goods (tea and coffee), services (tourist arrivals), production (vehicles produced, electricity generated), remittances, money supply (M2). - Soft data: purchasing managers’ sentiments (PMI). - Rationale for including specific series: - Coffee and tea exports: significant in Kenya’s agricultural production and trade; monitored by financial analysts. - Diaspora remittances: Kenya’s largest source of foreign exchange; support household consumption and investment. - Visitor arrivals: capture tourism contribution; key source markets include the USA, UK, India, Tanzania, and Uganda. - Electricity and vehicle production: important production variables; electricity generation reached approximately 13,000 GWh in 2024; installed capacity grew to 3.82 million kW in 2023; demand peaked at 2,362 MW in July 2025; vehicle production examples include 826 units in November 2024 and 978 units in October 2024; target production to 3,270 units by 2030. - Not included: Consumer Price Index is not included because the volatility of price indicators in Kenya is substantially higher than that of GDP growth. - Table 1: Selected economic indicators (excerpted key fields) - Stanbic Bank Kenya PMI — Frequency: Monthly — Start — End Dates: Jan 2014 — Dec 2025 — Lag (days): 5 — Unit: Index — Transformation: Level — Correlation (ρy,gdp): -0.01 — Variance share of Common Component: 0.000 — Source: S&P Global - Diaspora Remittances — Frequency: Monthly — Start — End Dates: Jan 2004 — Nov 2025 — Lag (days): 11 — Unit: NSA, Mil. USD — Transformation: y-o-y %∆ — Correlation (ρy,gdp): -0.02 — Variance share of Common Component: 0.210 — Source: CBK - Mobile Money — Frequency: Monthly — Start — End Dates: Mar 2007 — Nov 2025 — Lag (days): 17 — Unit: NSA, Bil. KES — Transformation: y-o-y %∆ — Correlation (ρy,gdp): 0.02 — Variance share of Common Component: 0.254 — Source: KNBS/CBK - Exports of Goods — Frequency: Monthly — Start — End Dates: Aug 1998 — Nov 2025 — Lag (days): 41 — Unit: NSA, Mil. KES — Transformation: y-o-y %∆ — Correlation (ρy,gdp): 0.19** — Variance share of Common Component: 0.299 — Source: CBK - Imports of Goods — Frequency: Monthly — Start — End Dates: Aug 1998 — Nov 2025 — Lag (days): 41 — Unit: NSA, Mil. KES — Transformation: y-o-y %∆ — Correlation (ρy,gdp): 0.36*** — Variance share of Common Component: 0.618 — Source: CBK - Exports of Coffee — Frequency: Monthly — Start — End Dates: Aug 1998 — Nov 2025 — Lag (days): 41 — Unit: NSA, Tonnes — Transformation: y-o-y %∆ — Correlation (ρy,gdp): -0.02 — Variance share of Common Component: 0.028 — Source: KNBS/CBK - Exports of Tea — Frequency: Monthly — Start — End Dates: Aug 1998 — Nov 2025 — Lag (days): 41 — Unit: NSA, Tonnes — Transformation: y-o-y %∆ — Correlation (ρy,gdp): -0.18** — Variance share of Common Component: 0.104 — Source: KNBS/CBK - Money Supply (M2) — Frequency: Monthly — Start — End Dates: Dec 1995 — Sep 2025 — Lag (days): 45 — Unit: NSA, Mil. KES — Transformation: y-o-y %∆ — Correlation (ρy,gdp): 0.11 — Variance share of Common Component: 0.018 — Source: CBK - Electricity Production — Frequency: Monthly — Start — End Dates: Feb 2006 — Nov 2025 — Lag (days): 48 — Unit: NSA, Mil. KWh — Transformation: y-o-y %∆ — Correlation (ρy,gdp): 0.30*** — Variance share of Common Component: 0.073 — Source: KPLC/KNBS - Visitor Arrivals (Persons) — Frequency: Monthly — Start — End Dates: Jan 2007 — Sep 2025 — Lag (days): 48 — Unit: Number — Transformation: Level — Correlation (ρy,gdp): 0.38*** — Variance share of Common Component: 0.069 — Source: KTB/KNBS - Vehicle Production — Frequency: Monthly — Start — End Dates: Jan 2003 — Nov 2025 — Lag (days): 48 — Unit: Number — Transformation: y-o-y %∆ — Correlation (ρy,gdp): 0.35*** — Variance share of Common Component: 0.055 — Source: KNBS - Real GDP — Frequency: Quarterly — Start — End Dates: 2009 Q1 — 2025 Q3 — Lag (days): 99 — Unit: NSA, Mil. KES — Transformation: y-o-y %∆ — Correlation (ρy,gdp): 1.00 — Variance share of Common Component: 0.368 — Source: KNBS - Notes: CBK—Central Bank of Kenya; KPLC—Kenya Power & Lighting Company; KNBS—Kenya National Bureau of Statistics; KTB—Kenya Tourism Board. ***, **, * indicate statistical significance of contemporaneous correlations at 1%, 5%, 10% level, respectively. The variance share of common component is estimated over the in-sample period spanning January 2010 and September 2021. ### 2.1.2 Transaction data - Rationale: use timely insights from non-traditional transaction data to better capture private demand; digital payments data were highly effective during the COVID-19 pandemic. - Novelty: to the authors’ knowledge, this is the first paper that uses digital payments data to nowcast GDP growth in Kenya. - Digital payments data: - Source: Total agent cash in cash out from the CBK (mobile payments data compiled by the CBK from digital payment platforms). - Publication timing: data published on the 17th of each month for the preceding month. - Reach and expansion (May 2025): - Active mobile money agents: 424,404 - Registered mobile money accounts: 85.62 millions - Number of mobile money transactions: 214.51 millions - Transaction value trends: - Value of mobile money transactions dropped by 21.8% y-o-y to 636.21 billion Kenya Shillings (about 3.7% of 2025 GDP) in February 2025. - Value in May 2025: 713.25 billion Kenya Shillings (about 4.2% of 2025 GDP) — still 1.2% lower than the same month in the previous year. - Interpretation: divergence between transaction counts (edging higher) and transaction value (declining) in H1 2025 suggests a shift toward low-value payments and a retreat in large-scale use, reflecting higher living costs and slowing household spending. - Market structure: - M-PESA market share: edged down to 91% in 2025 Q1 amid increasing competition (e.g., Airtel Money). - Financial access: over 85% of Kenyan adults have access to financial services mainly through digital platforms in 2025. - Historical behavior: mobile money sector plunged in March–April 2020, recovered by jumping 84% from its trough in April 2020 to February 2021, driven by Central Bank measures and shifts to contactless transactions. - Note: terms digital payment, digital money, and mobile money are used interchangeably. ### 2.1.3 Sampling frequencies and correlations - Frequency and lag: all indicators except quarterly GDP are sampled monthly and have a maximum delay of about one and a half months. - Real-time vintages: detailed historical real-time datasets for Kenya do not exist; the authors reconstructed pseudo data vintages starting from September 2021 to mimic actual data flow and evaluate the model in a pseudo-real-time setting. - Contemporaneous correlations: - Most contemporaneous correlations (ρy,gdp) in Table 1 are economically and statistically significant. - About half of the contemporaneous correlations exceed 0.2. - PMI and remittances have relatively weak absolute contemporaneous correlations but are timely (reporting lags of 5 and 11 days) and lead GDP growth. - Modeling implication: descriptive analysis supports application of a dynamic factor model to nowcast Kenya’s economic activity. - Note: Annex A provides additional details on the transactions. ### 2.2 Why nowcast Kenya’s GDP growth? #### 2.2.1 Main challenges to predict near-term GDP growth - Long delay in GDP releases: GDP data typically published with at least one quarter delay; in some LICs GDP is only available annually and subject to substantial revisions. - Limited data availability: only 11 monthly economic series publicly available for Kenya (Table 1) compared to 190,000 monthly US series in the St. Louis Fed database (accessed July 30, 2025). - Shorter historical data: shorter series make reliable model development more challenging. - Mixed frequencies and ragged edge: different frequencies (monthly, quarterly, annual), asynchronous release dates, missing observations create a sparse ragged-edge dataset structure complicating nowcasting. - Data discontinuities and series changes: more severe in developing and low-income countries where publication can be discontinued or frequencies changed. - Modeling approach to handle issues: treat mixed frequencies as temporal aggregation and periodically missing observations; handle ragged edge as missing data; use Kalman filter to interpolate and forecast missing observations while accommodating aggregation — full-system dynamic factor models jointly model target and predictors. - Advantage: Kalman filter allows computation of model-based surprises (news) for each release, enabling systematic evaluation of contributions from surveys, financial indicators, production, or labor market data. - Policy relevance: framework produces timely predictions and interpretable updates to help policymakers take informed decisions concerning stabilization measures. #### 2.2.2 Data releases and nowcasting - Timing implication: Kenya GDP publication lag of at least one quarter implies the need to estimate both current quarter GDP growth (nowcast) and previous quarter GDP growth (backcast). - Operational schedule: run a nowcast of current quarter growth and a backcast of previous quarter growth on the 7th and 21st day of each month. ### 3 Nowcasting framework #### 3.1 Dynamic factor model - Modeling choice: dynamic factor model (DFM) as in Giannone et al. (2008); a full forecast framework where target and predictors are endogenously and jointly determined and forecasted. - Model decomposition: each indicator decomposed into (1) common factors Ft and (2) idiosyncratic component εt. - Key assumptions: - Comovements summarized by a small set of unobserved common factors Ft. - Idiosyncratic components are orthogonal to each other and to common factors at each lead and lag (diagonal Θ and Σ). - Diagonal Θ and Σ preserve cross-orthogonality among idiosyncratic components; robustness to violations documented (Doz et al., 2012; Barigozzi and Luciani, 2024). - Model equations (state-space representation preserved as in source): - yt = ΛFt + εt - Ft = ΓFt−1 + ut, ut ∼ i.i.d. N(0, Ξ) - εt = Θεt−1 + ηt, ηt ∼ i.i.d. N(0, Σ) - State-space stacking and covariance structure as specified in the source. - Estimation strategy: expectation-maximization algorithm and the Kalman filter; initial estimation based on principal components (Doz et al., 2012). - Number of factors: run model with k ranging from 1 to 5; authors settle on a two-factor model (k = 2) as a parsimonious choice to avoid overfitting; note that a single-factor model is often sufficient in many nowcasting applications but results shown are a conservative lower bound of potential accuracy. *Source: wpiea2026032-source-pdf - 2.1 Data* ### 3.2 Mixed frequencies ### 3.2 Mixed frequencies ### Factor-based mixed-frequency aggregation - Nowcasting model is fitted to a dataset containing mixed-frequency (quarterly and monthly) indicators; a factor-based mixed-frequency aggregation bridges the frequency gap between monthly predictors and quarterly GDP growth. - Notation: the vector in Equation 1 is rewritten as y_t = [y^q_t,gdp; y^m_t]. - Quarterly GDP growth y^q_t,gdp is treated as partially observed at the monthly frequency, with its quarterly release assigned to the last month of the respective quarter. - Monthly unobserved GDP growth y^m_t,gdp is defined as the difference in log level GDP Y^m_t,gdp − Y^m_{t−1},gdp and features the same factor structure as other monthly indicators: - y^m_t,gdp = λ_gdp f_t + ε_t,gdp. (7) - ε_t,gdp = ρ_gdp ε_{t−1},gdp + η_t,gdp. (8) - Aggregation mapping monthly to quarterly year-on-year GDP growth (partially observed every third month): - y^q_t,gdp = y^m_t,gdp + y^m_{t−1},gdp + y^m_{t−2},gdp - = λ_gdp (f_t + f_{t−1} + f_{t−2}) + ε_t,gdp + ε_{t−1},gdp + ε_{t−2},gdp. (9) - Implication: quarterly variables load equally on current and two lagged values of the unobserved monthly factor; monthly factor series are extracted with the Kalman filter and plugged into Equation 9 to obtain the nowcast for quarterly GDP growth. - Note: the same approach could be applied when GDP growth is available at even lower frequencies (e.g., annual or biannual). ### Estimation: expectation-maximization algorithm - Estimation uses an expectation-maximization (EM) approach (Watson and Engle, 1983; Doz et al., 2012) adapted for ragged-edge nowcasting panels (Baňbura and Modugno, 2014) to handle arbitrary patterns of missing data. - Definitions and initialization: - T_υ denotes the time index of the most recent observation in the information set Ω_υ. - Log-likelihood expressed as L(y,F,Ψ) with y = (y_1,...,y_{T_υ}), F = (F_1,...,F_{T_υ}), and Ψ = (Λ,Γ,Θ,Σ,Ξ). - EM initialized at iteration ι = 0: factors F^{(0)}_t extracted by Principal Component Analysis (PCA); parameters Ψ^{(0)} estimated by ordinary least squares (OLS). - Iterative steps until convergence: - E-step: given current estimates ˆΨ^{(ι)}, compute expected log-likelihood L(Ψ,Ψ^{(ι)}) = E_{Ψ^{(ι)}}[L(y,F,Ψ)|Ω_υ] and extract factors using the Kalman filter and smoother. - M-step: re-estimate parameters by maximizing the expected log-likelihood Ψ^{(ι+1)} = arg max_Ψ L(Ψ,Ψ^{(ι)}). - Practical notes: - Simplicity of steps makes algorithm computationally viable without size limitations; each step increases likelihood but convergence to a global maximum is not guaranteed (local maxima possible). - PCA typically provides a good starting point close to the MLE (Doz et al., 2012). - Bayesian alternatives exist (D’Agostino et al., 2016; Antolín-Díaz et al., 2024) using priors and Gibbs sampling: (1) draw missing data and latent factors conditional on parameters via simulation smoother; (2) draw parameters conditional on full data and latent factors from posterior distributions. - Ancillary references: EM details and state-space representation in Annex Section C and Section B of the Annex. ### Real-time data flow and news decomposition - Objective: establish framework to utilize flow of real-time data releases for nowcasting Kenya’s GDP growth; model produces nowcasts and forecasts for all variables, enabling extraction of news (unexpected component) from released data. - News-driven nowcast updating: - Consider two consecutive data vintages Ω_v and Ω_{v+1} (variables released in months v and v+1). Ω_v ⊂ Ω_{v+1} and Ω_{v+1}\Ω_v = y^t_{j,n_j}, for j = 1,...,J_{v+1}. (10) - Nowcast revision identity (conditional expectation as orthogonal projection): - E[y^n_{t,n}|Ω_{v+1}] = E[y^n_{t,n}|Ω_v] + E[y^n_{t,n}|I_{v+1}]. (11) - I_{v+1} = [I_{v+1,1} ... I_{v+1,J_{v+1}}]' with I_{v+1,j} = y^t_{j,n_j} − E[y^t_{j,n_j}|Ω_v]; I_{v+1} is the “news” (orthogonal to Ω_v). - Nowcast revision expressed as weighted sum of news: - E[y^n_{t,n}|Ω_{v+1}] − E[y^n_{t,n}|Ω_v] = Σ_{j∈J_{v+1}} δ_{v+1,j} (y^t_{j,n_j} − E[y^t_{j,n_j}|Ω_v]). (12) - J_{v+1} is the set of newly released variables at time v+1; δ_{v+1,j} are associated weights. - Interpretation: - Revision magnitude depends on both news magnitude and relevance (weight δ_{v+1,j}). - Equation 12 generalizes the Kalman filter update to asynchronous arrivals; filter uncertainty for y^n_{t,n} diminishes with new data releases and can be similarly decomposed. - When multiple variables are released simultaneously, forecast revision can be decomposed into contributions from each series or group. - A full prediction model predicting all variables in real time is necessary to construct transparent news decomposition; partial models do not provide complete model-based impacts of news and change specification with data availability. - Practical complications: - Formula in Equation 12 assumes fixed model parameters while Ω_v expands only via new releases; in practice, historical data are often revised, so model parameters are reestimated iteratively (Section 3.3 and Annex Section C). - Kalman smoother output provides the news I_{v+1} and expectations needed to compute weights δ_{v+1,j} (Baňbura and Modugno, 2010; Baňbura et al., 2011, 2013). - MIDAS is noted as a computationally convenient partial prediction approximation (Bai et al., 2013), but the Kalman filter-based estimation used here is computationally efficient and viable with very large datasets (Doz et al., 2012); partial-model drawbacks relative to full-system DFM are highlighted. ### Empirical implementation and model selection (summary points linking mixed frequencies to empirical practice) - Model choice and sample: - Two-factor DFM chosen for Kenya based on empirical findings: single-factor models fail to capture the COVID-19 recession and recovery dynamics; multi-factor models (≥2 factors) capture pandemic dynamics better and models with more than two factors are virtually indistinguishable. - Estimated from January 2010 to September 2021. - Two-factor model’s common component closely tracks GDP growth and sectoral indicators, capturing normal-cycle fluctuations and large swings (sharp decline in 2020 and recovery in 2021). - Annexes E–I present DFM results for one- and three-factor models; nowcasts from 4- and 5-factor models are available from the authors upon request. - Real-time forecasting exercise design and timing: - Simulated forecaster runs model in real time every two weeks beginning on September 21, 2021. - Model parameters recursively re-estimated at the start of each quarter using the most recent 12 years of data; parameters are fixed within the quarter for the nowcast prediction exercise. - Nowcasts generated solely from variables that would have been released by that time following the stylized data availability calendar in Section 3.4. - Pseudoreal-time period: sequential forecasts from September 2021 to March 2024 (using constructed pseudo-vintages mimicking real-time data flow). - Truereal-time forecast extends from April 2024 to December 2025 (using real-time vintages for series listed in Table 1). - Prediction frequency in each quarter: - Six nowcasts of the current-quarter GDP growth (two nowcasts on the 7th and 21st day of each of the three months in the quarter). - Six backcasts of the previous-quarter GDP growth to mimic historical real-time predictions. - The nowcast revision visualization (Figure 4) constructs a weekly progression with black diamonds representing weekly nowcasts and colored bars showing contributions of the news component from all 12 variables; black triangle indicates the impact per Equation 12. - Pseudoreal-time quarterly predictions constructed for 2021 Q3 to 2024 Q1; truereal-time forecasts span 2024 Q2 to 2025 Q4. Final predictions for example quarters shown are 2024 Q3 — 2025 Q4; nowcasts for other quarters are available in Annex F. *Source: wpiea2026032-source-pdf - 3.2–4.1* ### 2024. We produce bimonthly updates of nowcasts, forecasts, and backcasts, resulting in ### wpiea2026032-source-pdf - 2024. We produce bimonthly updates of nowcasts, forecasts, and backcasts, resulting in ### Nowcast updates and timeline (2024 Q3 — 2024 Q4) - Nowcast schedule: 12 predictions for each quarter; nowcast covers the three months of the reference quarter plus the three months following it, up to the GDP publication date. - First update for the 2024 Q3 nowcast: July 7, coinciding with PMI, M2, and first-quarter GDP growth figures. - July 7–July 21 revisions: - Nowcast fell from 4.06% to 3.57% between updates on July 7 and July 21. - Main negative drivers in July: second-quarter growth rate, PMI, and M2. - Broad-based downward revision except visitor arrivals and tea exports. - Largest downward adjustments: imports, exports, and mobile money (digital money); followed by electricity production and remittances. - August–October 2024 pattern: - GDP growth decline begins in July, stabilizes in August, picks up in September and October driven by positive readings in imports, exports, vehicle production, visitor arrivals, and remittances. - Subsequent downward revisions as most macro indicators are released, reflecting weak digital money, trade, and vehicle production. - Backcasting for 2024 Q3 begins October 2024 (start of Q4); backcasted growth stays around 3.87% through end-December 2024. - Predicted GDP growth for 2024 Q3: 4.65%; actual recorded growth for 2024 Q3: 4.2%. ### Drivers and variable contributions (key findings) - Trade variables are the most significant contributors to nowcast revisions: - Exports and imports exert the largest downward influence in July 2024. - Imports drive the upward revision observed in August and September 2024. - Digital money (mobile money) is crucial in July and August revisions, with a particularly notable negative contribution in August 2024. - Timely variables provide early signals: - PMI acts as a leading indicator early in the nowcast quarter, with little to no effects toward the end. - Useful to include both timeliest indicators (signaling direction) and variables with strong comovement. - Vehicle production increase in September 2024 led to an upward revision, later revised downward in October 2024. ### Real-time revisions across 2024 Q4 – 2025 Q4 (Panels B–F summary) - 2024 Q4 nowcast: - Downward revisions continued until January 2025, driven by disappointing trade and digital money readings, then stabilized through March 2025. - 2024 Q4 GDP data release on May 6, 2025 led to a large upward revision from 3.12% to 4.00%. - 2025 Q2 nowcast: - Gradual upward revision from 3.87% in April to 5.18% in September 2025, approaching the actual growth rate of 5.0%. - Primary drivers of the upward adjustment: imports, followed by vehicle production; digital money provided an initial substantial contribution in April 2025. - 2025 Q3 nowcast: - Final prediction for 2025 Q3: 4.74%; actual published growth by KNBS on December 31, 2025: 4.90%. - 2025 Q4 nowcast: - Forecast edged down from 4.61% on October 7, 2025 to 4.16% on December 21, 2025, primarily due to weak digital money readings, before recovering to 4.50% in January 2026. ### Impact of macroeconomic data releases (Figure 5 insights) - Four key insights from average absolute weekly impact (real time, 2022–2025 sample): - The publication of previous-quarter GDP growth significantly affects nowcast updates for both nowcasts and backcasts. - Most important sectors: trade (exports and imports), production (electricity and vehicle), and consumption (digital money). - Positive association between imports and GDP growth: imports act as catalysts via technology-embedded capital goods and intermediate inputs, supporting productivity, industrialization, and export competitiveness. - Mobile/digital money has substantial contribution: facilitates formal and informal sector activity, boosts financial inclusion, consumption smoothing, risk-sharing, and poverty alleviation. - Timing of useful information: - Digital money figures published 15 to 17 days after month-end. - PMI is the timeliest variable; contributes mainly at the beginning of the sample with influence fading over time. - The most useful information arrives in the middle of the nowcasting period (bell-shaped pattern); early and late surprises have less impact. ### Predictive performance and nowcast errors (out-of-sample evaluation) - Overall model behavior: - Model tracks actual GDP growth closely over time, capturing 2022 slowdown and 2023 rebound. - Nowcast accuracy improves as more data arrive closer to GDP publication. - Specific quarterly examples: - 2024 Q1 nowcasts (January–March 2024): revised up from 5.06% to 6.40% (six nowcasts). - 2024 Q1 backcasts (April–June 2024): remain around 6.30% until official GDP publication in early July 2024 at 4.94% (actual lower than predicted). - 2024 Q2 nowcasts (April–June 2024): values ranged from 5.78% to 4.83%; backcasts updated to values between 3.47% and 3.89% until 2024 Q2 GDP growth of 4.6% published on October 2, 2024. - 2025 Q1 nowcasts: revised downward from 4.16% to 3.12%, driven by negative news in mobile money figures; 2024 Q4 GDP data release partly offset negative effects. - 2025 Q2 and Q3 predictions hovered around 5%, in line with realized figures. - Nowcast errors: - Defined as Error_t = ˆE[y_t,gdp | Ω_T] − y_t,gdp for 0 ≤ t ≤ T. - Nowcast errors shrink toward zero as time approaches official GDP data release (errors computed across 15 quarters from 2021Q4 to 2025Q3). - Dots at horizon one show revision error as of one day before GDP release; averaged across quarters to produce error profiles by weeks until publication. ### Annual predictions and benchmarking (calendar year assessments) - Annual nowcasts constructed by aggregating quarterly forecasts updated bimonthly. - Nowcast for 2024 (Panel A, Figure 8): - Nowcasts decreased from 5.60% to 5.07% in the first half of 2024 because of negative trade developments, then rebounded and stabilized to 4.72% in 2025 Q1, matching actual GDP growth before official release. - Nowcasts were significantly lower than other institutional predictions in 2024 Q1 but gradually converged toward CBK and WEO forecasts, surpassing them slightly in 2025 Q2 by 2 and 21 basis points, respectively. - Nowcast for 2025 (Panel B, Figure 8): - As of the reporting date, 10 data points still missing from required 33. - On December 7, 2025 the model predicted an increase in growth to 4.72% in 2025, up from 4.37% in January 2025. - Nowcasts for 2025 were substantially lower than institutional predictions in Q1 but converged toward WEO projections in Q3 and Q4. - Backtesting annual GDP growth (Figure 9): - 2022 backcasts: dropped from 5.94% to 4.94% in 2022 Q3, then gradually to 4.87% in 2023 Q1, close to realized GDP growth of 4.90%; CBK and WEO forecasts remained higher (~5.6% and 5.37% in 2023 Q1). - 2023 backcasts: rose from 3.44% in 2023 Q1 to 5.72% in 2024 Q1; CBK and WEO forecasts ranged between 5% and 6.1%. - Interpretation: CBK and WEO forecasts display stronger inertia; the DFM nowcasts are more flexible and track actual GDP growth more closely as information accumulates. _Italic: Source content unit "wpiea2026032-source-pdf - 2024. We produce bimonthly updates of nowcasts, forecasts, and backcasts, resulting in" (excerpt)._ ### Section 3. GDP release dates follow the official calendar. Growth is measured as the year-over-year cha ### Section 3. GDP release dates follow the official calendar. Growth is measured as the year-over-year change. ### Nowcasting framework and data flow - Annual nowcasts are constructed by aggregating quarterly forecasts, which are updated on a bimonthly basis. - The modeling framework is a dynamic factor model tailored to Kenya’s economic conditions that: - Uses monthly indicators of activity in key Kenyan economic sectors. - Employs Kalman filtering techniques to produce early estimates of quarterly GDP growth. - Decomposes each data release into a predictable part and a news component to show how incoming information shapes nowcast updates. - CBK = Forecasts from the Central Bank of Kenya; WEO = Forecasts from the IMF World Economic Outlook. ### Empirical context and data quality - Kenya has been publishing quarterly GDP since 2009. - The reporting delay typically exceeds three months after the end of the quarter, complicating timely policy implementation in a context where data are published infrequently and with significant delay. - Nowcasting addresses the infrequent and delayed publication of activity measures by producing continuous near-term GDP growth estimates as new macroeconomic data arrive. ### Main findings and model performance - The single-factor (common component) framework, well established for advanced economies (AEs), can be meaningfully extended to a low-income country (LIC) context such as Kenya. - The nowcasting model shows strong and robust performance: - Both quarterly and annual GDP predictions from the nowcasting framework perform well out-of-sample compared to alternative forecasts by the CBK and the IMF. - Estimation results identify specific data releases that have been key for nowcast updates since 2024: - Electricity production - Vehicle production - Digital money - Exports of goods - Imports of goods ### Policy relevance and practical implications - The model offers a practical tool to continuously update near-term GDP growth predictions, informing timely stabilization policies and helping calibrate development policies with sound data-driven insights. - The framework’s ability to decompose news versus predictable components helps policymakers understand which incoming data drive revisions and why. - The approach is flexible and can seamlessly include new predictors as the set of high-frequency variables and non-traditional indicators grows, paving the way for applying this nowcasting approach to other LICs with infrequent or annual GDP series. *Source: wpiea2026032-source-pdf - Section 3. GDP release dates follow the official calendar. Growth is measured as the year-over-year change.* ### 237. Elsevier. ### 237. Elsevier. ### Major literature cited - Bánbura, M., Giannone, D., and Reichlin, L. (2011). Nowcasting. In Clements, M. P. and Hendry, D. F., editors, The Oxford Handbook of Economic Forecasting, pages 193–224. Oxford University Press. - Barigozzi, M. and Luciani, M. (2024). Quasi maximum likelihood estimation and inference of large approximate dynamic factor models via the EM algorithm. Finance and Economics Discussion Series 2024-086. Washington: Board of Governors of the Federal Reserve System. - Additional referenced works include methodological and applied contributions on nowcasting, dynamic factor models, big data in macroeconomics, payment-system data, mobile money, and country-specific nowcasting exercises (selected authors: Bokan, Burns & Mitchell, Caruso, Cascaldi-Garcia, Cerra & Saxena, Chetty et al., Dahlhaus et al., D’Amuri & Marcucci, Doz & Fuleky, Doz et al., Giannone et al., Ley & Misch, Luciani, Mishra et al., Modugno, Stock & Watson, Suri et al., Varian, Watson & Engle, Yao et al.). ### Annex A — Time series of selected economic indicators: data definitions, timing, units, and notable datapoints - Stanbic Bank Kenya PMI - Series initiated in January 2014. - Constructed from five components: new orders, production, employment, supplier deliveries, inventories. - Released monthly, within the first 5 days of the month for the previous month. - Enters the model at levels. - Data updated on the 7th of each month. - Diaspora Remittances - Released by the Central Bank of Kenya (CBK) on the 11th day of each month for the preceding month. - Presented in millions of USD and seasonally adjusted. - In 2024, remittances surged by 18 percent to US$4.9 billion. - Anticipated to contribute an average of 3.5 percent to GDP from 2025 to 2029 (Economist Intelligence Unit, July 21st, 2025). - Enters the nowcasting model using y-o-y growth rates and updated on the 21th of each month. - Exports and Imports of Goods - Collected and reported by the CBK within the first 41 days of each month for the previous month. - Presented in hundreds of millions of Kenyan Shillings (KES); not seasonally adjusted. - Follow IMTS definitions. - Incorporated into the model using monthly y-o-y growth rates. - Updated on the 21th day of the month for the previous two months. - Exports of Coffee - Released by KNBS and CBK 41 days from the previous publication. - Reported in Tonnes; not seasonally adjusted; follow IMTS definitions. - Entered into the model using monthly y-o-y growth rates. - Updated on the 21th day of the month for the previous two months. - Exports of Tea - Published by KNBS and CBK 41 days from the last publication. - Reported in Tonnes; not seasonally adjusted; follow IMTS definitions. - Included in the model using monthly y-o-y growth rates. - Updated on the 21th day of the month for the previous two months. - Money Supply M2 - Released by the CBK within 45 days of the previous publication; in the monthly Financial Statistics Bulletin. - Official release unit: KES Million. - Entered into the model using y-o-y growth rates. - Updated on the 7th day of the month for the previous two months. - Electricity Production - Released by Kenya Power & Lighting Company (KPLC) and the KNBS 48 days following the previous publication. - Official release unit: million kWh. - Entered into the model using monthly y-o-y growth rates. - Data for January–April indicates significant increases; updates occur on the 21th of each month for the previous two months. - Visitor Arrivals - Published by Kenya Tourism Board and KNBS within 48 days from the previous publication. - Key source markets: USA, UK, India, Tanzania, Uganda. - Inflows from international tourism exceeded pre-coronavirus levels in 2023; visitor arrivals were flat in the first quarter, with a 9.7% year-on-year increase in April. - Enters the model at level; updates on the 21th of each month for the past two months. - Vehicle Production - Released by the KNBS 48 days after the previous release. - Incorporated into the model using y-o-y growth rates. - Updates occur on the 21st of each month for the previous two months. - Data for January–April indicates strong increases in vehicle production. - Mobile Payments (Total agent cash in cash out) - Compiled by the CBK from digital payment platforms and published until February 2025; most recent months (March to May 2025) released with a 17-day delay in KNBS’ May 2025 report (published late July 2025). - Updated on the 21st of each month to account for corrections and checks. - Total agent cash in cash out: - Fell by 21.8% y-o-y to 636.21 billion Kenya Shillings in February 2025. - Fell by 1.2% y-o-y to 713.25 billion Kenya Shillings in May 2025. - Number of transactions increased even as transaction value fell—suggesting a shift toward low-value payments in H1 2025. - World Bank Findex: 82.3% of adults engaged with mobile money accounts in 2024; rural adoption at 77%. - Sector experienced a sharp decline in March–April 2020, followed by a robust rebound surpassing pre-pandemic levels; policy measures included fee waivers, raised wallet and transaction limits, and promotion of digital payments. - Mobile money services expanded into credit, savings, bill payments; supported e-commerce and remote work adoption. ### Section B — Bridging the frequency gap with factors: aggregation formula and state-space expression - Relationship between quarterly y-o-y GDP growth y^q_{t,gdp} and monthly y-o-y GDP growth y^m_{t,gdp}: - y^q_{t,gdp} = Y^q_{t,gdp} − Y^q_{t−12,gdp} = (1 − L^{12})Y^q_{t,gdp} ≈ (1 − L^{12})(1 + L + L^2)Y^m_{t,gdp} = (1 + L + L^2)(1 − L^{12})Y^m_{t,gdp} = (1 + L + L^2)y^m_{t,gdp} = y^m_{t,gdp} + y^m_{t−1,gdp} + y^m_{t−2,gdp}, (13) - Quarterly GDP growth re-expressed: - y^q_{t,gdp} = λ_gdp(f_t + f_{t−1} + f_{t−2}) + ε_{t,gdp} + ε_{t−1,gdp} + ε_{t−2,gdp}, (14) - State-space form (variables and loading structure shown in Equation 15), where Λ_m is an ((M−1)×1) vector of factor loadings for monthly predictors, 0 is an ((M−1)×1) vector of zeros, and I_m is an ((M−1)×1) vector of ones. (See Equation 15 for the stacked representation.) ### Section C — Details on the expectation-maximization (EM) algorithm for ragged-edge datasets - M-step re-estimation formulas in the (ι + 1)th iteration: - ˆΛ^{(ι+1)} = [sum_{t=1}^{T_υ} E_{Ψ^{(ι)}}(y_t F'_t | Ω_υ)] [sum_{t=1}^{T_υ} E_{Ψ^{(ι)}}(F_t F'_t | Ω_υ)]^{−1}, (16) - ˆΓ^{(ι+1)} = [sum_{t=1}^{T_υ} E_{Ψ^{(ι)}}(F_t F'_{t−1} | Ω_υ)] [sum_{t=1}^{T_υ} E_{Ψ^{(ι)}}(F_{t−1} F'_{t−1} | Ω_υ)]^{−1}, (17) - ˆΘ^{(ι+1)} = [sum_{t=1}^{T_υ} E_{Ψ^{(ι)}}(y_t y'_{t−1} | Ω_υ) − ˆΛ^{(ι+1)} sum_{t=1}^{T_υ} E_{Ψ^{(ι)}}(F_t y'_{t−1} | Ω_υ)] × [sum_{t=1}^{T_υ} E_{Ψ^{(ι)}}(y_{t−1} y'_{t−1} | Ω_υ) − ˆΛ^{(ι+1)} sum_{t=1}^{T_υ} E_{Ψ^{(ι)}}(F_{t−1} y'_{t−1} | Ω_υ)]^{−1}, (18) - ˆΣ^{(ι+1)} = diag{ T^{−1} sum_{t=1}^{T_υ} E_{Ψ^{(ι)}}(y_t y'_t | Ω_υ) − ˆΛ^{(ι+1)} T^{−1} sum_{t=1}^{T_υ} E_{Ψ^{(ι)}}(F_t y'_t | Ω_υ) }, (19) - ˆΞ^{(ι+1)} = T^{−1} sum_{t=1}^{T_υ} E_{Ψ^{(ι)}}(F_t F'_t | Ω_υ) − ˆΓ^{(ι+1)} T^{−1} sum_{t=1}^{T_υ} E_{Ψ^{(ι)}}(F_{t−1} F'_t | Ω_υ). (20) - Convergence criterion: increase in likelihood between consecutive iterations falls below an arbitrarily small threshold. - Handling missing observations in y_t (ragged-edge): - Represent y_t = (I_{M×M} − S_t) y_t + S_t y_t, where S_t is a diagonal selection matrix with ones for non-missing observations and zeros otherwise. (22) - Modified EM expressions for ragged-edge datasets: - vec(ˆΛ^{(ι+1)}) = [sum_{t=1}^{T_υ} E_{Ψ^{(ι)}}(F_t F'_t | Ω_υ) ⊗ S_t]^{−1} vec{sum_{t=1}^{T_υ} S_t y_t E_{Ψ^{(ι)}}(F'_t | Ω_υ)}, (23) - ˆΣ^{(ι+1)} = diag{ T^{−1} sum_{t=1}^{T_υ} [ S_t y_t y'_t S'_t − S_t y_t E_{Ψ^{(ι)}}(F'_t | Ω_υ) ˆΛ^{(ι+1)'} S'_t − S_t ˆΛ^{(ι+1)} E_{Ψ^{(ι)}}(F_t | Ω_υ) y'_t S'_t + S_t ˆΛ^{(ι+1)} E_{Ψ^{(ι)}}(F_t F'_t | Ω_υ) ˆΛ^{(ι+1)'} S'_t + (I_{M×M} − S_t) ˆΣ^{(ι)} (I_{M×M} − S_t) ] }, (24) ### Figures and empirical diagnostics summary (selected highlights) - News impacts (Figure D11) - Average absolute impacts of macroeconomic variables on nowcasts computed in real time and averaged over 2021 Q1 to 2025 Q4. - Variables include PMI, REM, ARR, ELEC, VEH, M2, EXP, IMP, COF, TEA, MCASH, GDP. - Panels show comparisons for 1-factor and 3-factor models across bimonthly updates prior to the first official GDP release. - Nowcast errors (Figure E12) - Forecast errors computed in out-of-sample exercise from 2021Q4 to 2025Q3 using the nowcasting model (Section 3.1). - Dots represent revision errors for all 14 quarters; x-axis denotes weeks until GDP publication (24 to 2 weeks). - Vertical black dotted line separates nowcast period (24 to 14 weeks before GDP publication) from backcast period (12 to 2 weeks). - Real-time predictions and news decomposition (Figures F13–H20, G15–G17) - Quarterly nowcasts and contributions of predictors shown for 2022 Q1–2025 Q4 across 1-factor, 2-factor, and 3-factor models. - Black solid line: quarterly nowcasts of GDP growth; red hollow diamond: published actual GDP growth. - Black triangle: change in two consecutive quarterly nowcasts; bars: contribution of each predictor (PMI, REM, ARR, ELEC, VEH, M2, EXP, IMP, COF, TEA, MCASH) to nowcast updates. - Annual nowcasts and backcasts (Figure I21) - Nowcasting and backcasting annualized quarterly GDP growth shown for backcasts of 2022 and 2023, and nowcasts for 2024 and 2025 using the 1-factor model. - Solid line depicts evolution of annualized quarterly GDP growth backcasts using the model. *Italicized source: wpiea2026032-source-pdf - 237. Elsevier.* ### Section 3. GDP release dates follow the official calendar. Growth is measured as the year-over-year cha ### Section 3. GDP release dates follow the official calendar. Growth is measured as the year-over-year change. ### Methodology and Timing - Growth is measured as the year-over-year change. - Annual nowcasts are constructed by aggregating quarterly forecasts, which are updated on a bimonthly basis. - The solid line in Figure I22 depicts the evolution of annualized quarterly GDP growth backcasts using the model described in this section. ### Data sources and forecast comparisons - CBK = Forecasts from the Central Bank of Kenya; WEO = Forecasts from the IMF World Economic Outlook. - CBK forecasts from 2025 Q3 and Q4 are not publicly available. ### Nowcasting and backcasting results (Figure I22) - Figure I22 presents model backcasts and nowcasts of annual GDP growth (3-factor model) with the vertical axis labeled "Annual GDP growth in %" showing tick marks at 4, 4.5, 5, 5.5, 6, and 6.5. - Panels in Figure I22: - (A) Backcast for 2022: timeline from 2022-01 to 2023-10 showing Annualized, Nowcast, WEO, CBK, and Realized series. - (B) Backcast for 2023: timeline from 2023-01 to 2024-10 showing Annualized, Nowcast, WEO, CBK, and Realized series. - (C) Nowcast for 2024: timeline from 2024-01 to 2025-10 showing Annualized, Nowcast, WEO, CBK, and Realized series. - (D) Nowcast for 2025: timeline from 2025-01 to 2026-10 showing Annualized, Nowcast, WEO, CBK, and Realized series. ### Key statistics and timeline details - Axis tick values explicitly shown in the figure: 4; 4.5; 5; 5.5; 6; 6.5 (Annual GDP growth in %). - Figure time ranges by panel: - Panel (A): 2022-01, 2022-04, 2022-07, 2022-10, 2023-01, 2023-04, 2023-07, 2023-10. - Panel (B): 2023-01, 2023-04, 2023-07, 2023-10, 2024-01, 2024-04, 2024-07, 2024-10. - Panel (C): 2024-01, 2024-04, 2024-07, 2024-10, 2025-01, 2025-04, 2025-07, 2025-10. - Panel (D): 2025-01, 2025-04, 2025-07, 2025-10, 2026-01, 2026-04, 2026-07, 2026-10. *Nowcasting GDP Growth for Kenya — Working Paper No. WP/2026/032* --- _Source: https://www.imf.org/-/media/files/publications/wp/2026/english/wpiea2026032-source-pdf.pdf_