In an increasingly digital environment where data and advanced analytics challenge traditional economic modeling, the Bank of England is applying a fusion of machine learning (ML) with economic theory to better understand complex phenomena like inflation.

Two working papers shared by the Bank of England spotlight this development.

Authored by experienced researches and industry professionals, these studies—aim to address or tackle the “black box” opacity of ML head-on.

By embedding economic principles into algorithms, they aim for sharper policy tools for central banks navigating volatile post-pandemic economies.

The first paper lays the conceptual groundwork, arguing that raw ML’s predictive prowess often sacrifices interpretability, rendering it ill-suited for policy scrutiny.

Buckmann and Potjagailo introduce structured ML techniques, notably Block-Additive Models (BAMs) and theory-consistent monotonicity constraints, to infuse economic logic without diluting performance.

BAMs partition predictors into “blocks” based on economic rationale—say, grouping supply-chain variables or labor market indicators—allowing non-linear interactions within blocks while enforcing additivity across them.

This yields clear, attributable contributions: one block might reveal how oil prices nonlinearly spike inflation, while another isolates wage pressures.

Monotonicity constraints further align models with theory, ensuring outputs respect established relationships, like inflation rising monotonically with demand but falling with supply gluts.

Applied to simulated and real datasets, these methods retain ML’s forecasting edge—outperforming linear regressions—while delivering “economically meaningful narratives.”

The implications are profound for central banking: policymakers can now probe “what-if” scenarios, such as the inflationary fallout from a trade war, with models that echo Phillips curve dynamics rather than inscrutable neural networks.

Building on this foundation, the second paper operationalizes these ideas in a tailored inflation model: the Blockwise Boosted Inflation Model (BBIM).

Here, the trio deploys boosted trees—a gradient-boosting ML variant—within a block-structured framework inspired by the open-economy hybrid Phillips curve.

Inflation is dissected into demand and supply blocks, with monotonicity ensuring intuitive links: tighter labor markets unambiguously stoke price pressures, while excess capacity cools them.

Calibrated on UK Consumer Prices Index (CPI) data from 1997 to 2024, BBIM uncovers non-linear drivers that linear models miss.

The recent inflation surge, peaking at 11.1% in October 2022, emerges not as a uniform wage-price spiral but as a supply-shock cascade.

“The model shows that the recent surge has been driven mainly by global supply shocks transmitted through supply chains,” the authors note, pinpointing disruptions like the Ukraine war’s energy ripple effects.

Labor markets reveal an “L-shaped” Phillips curve: slack broadly suppresses inflation during recessions, but in booms, tightness unleashes asymmetric upward bursts, amplified by post-COVID hiring frenzies.

Household inflation expectations add another layer of nuance.

Short-term views, swayed by headline spikes, exhibited persistent non-linear effects, temporarily elevating trend inflation by 0.5-1 percentage points.

Long-term anchors, however, held firm at the Bank’s 2% target, underscoring the resilience of credibility.

Out-of-sample forecasts from 2022 onward beat benchmarks, including unstructured ML and vector autoregressions, by up to 20% in mean absolute error.

These papers collectively seemingly indicate a paradigm shift.

Traditional econometrics excels in transparency but falters on big data’s nonlinearities; pure ML thrives on patterns yet baffles decision-makers.

The Bank of England’s hybrid—structured, boosted, and economically tethered—aims to bring together the best of both.

For inflation forecasting, BBIM equips the Monetary Policy Committee with granular insights: supply shocks demand fiscal buffers, while labor non-linearities signal timely rate hikes.

Broader impacts extend to global finance.

As climate shocks and geopolitical fractures intensify, interpretable ML could forecast spillovers in emerging markets or even try to dissect crypto’s deflationary proposition.

Yet challenges linger: scaling these models to real-time data requires computational heft, and over-reliance on UK-specific blocks potentially risks parochialism.

Ultimately, these works affirm ML’s maturation in economics.