×
Macro Research · Quantitative Inflation Strategy
U.S. Inflation Outlook & Scenario Analysis
Machine-learning based probabilistic forecasts — forward view from April 2026
Prepared: April 7, 2026
Horizon: 12 months (through April 2027)
Model: InflationTransformer v1.0
Confidential — Internal Use Only
Video Walkthrough

Iran. Tariffs. Oil. — A walkthrough of the InflationTransformer model, 12 macro scenarios, and risk probabilities for April 2026. Runtime ~8 min.

Executive Summary

CPI inflation currently stands at approximately 2.2% YoY (April 2026), down from a peak of ~9% in mid-2022. Our proprietary transformer-based model projects that under a no-new-shock baseline, headline CPI will remain near 2.7–2.8% through April 2027 — above the Fed's 2% target but well within a manageable range.

The critical risk to this baseline is policy-driven supply shocks. Even moderate tariff escalation raises the 12-month CPI median to 5.0% with a 50% probability of exceeding 5%. A broad stagflation scenario (combined supply, wage, and energy shocks) produces an 8.6% median at M+12 — re-entering 2022 crisis territory. Only a severe demand contraction returns CPI below the 2% threshold.

Current CPI YoY
2.2%
April 2026
Baseline M+12
2.74%
No new shocks (oil flat)
Tariff Severe M+12
6.2%
P(CPI>5%) = 82%
Stagflation M+12
8.6%
P(CPI>5%) = 90%
Key strategic implication: The dispersion across scenarios is exceptionally wide — 0.3% to 8.6% at M+12. This is not a forecast uncertainty problem; it is a policy path uncertainty problem. The model's baseline is benign, but the upside tail (tariff + supply shocks) is severe and non-linear beyond month 6.

Research Approach & Data

This report is produced by InflationTransformer, a probabilistic deep learning model trained on 26 years of monthly U.S. macroeconomic data (1998–2024). The model produces a full probability distribution — not a single point forecast — for CPI over a 12-month horizon, conditioned on user-specified macro scenarios.

Training Data

312 monthly observations, 1998–2024. 15 macro variables sourced from FRED and Yahoo Finance. Covers 3 Fed tightening cycles, GFC, COVID spike, and 2021–2022 re-acceleration.

Model Architecture

Multi-scale patch transformer with Gaussian Mixture output (3 components). Takes 48 months of history as input; produces 12-month probabilistic forecast simultaneously. Oil/energy scenario is injected as a conditioning signal.

Scenario Engine

12 pre-defined macro scenarios covering oil price paths, trade policy shocks, stagflation, demand contraction, and commodity supercycles. Each scenario applies shocks to multiple variables and runs 500 Monte Carlo draws to estimate full distributions.

Structural Overlay

Literature-calibrated pass-through multipliers applied post-model: Oil (×0.05), PPI (×0.10), Import Prices (×0.08), Wages (×0.20). Ensures macroeconomic credibility of scenario paths beyond what training data alone covers.

Key Macro Variables

CPI headline & core, PCE, Oil WTI, PPI, Import Prices, Wages (ATW/ECI), M2 money supply, 10Y–2Y yield spread, USD index, University of Michigan expectations, Leading Indicators, Fed Funds Rate.

Validation

Rolling 1-step-ahead backtest on 2021–2026 data (never seen in training). OOS RMSE: 3.1pp full period (dominated by COVID spike); post-peak RMSE ~0.6pp. 80% prediction intervals cover 79% of actual observations — well-calibrated.

Model Performance — Out-of-Sample Backtest

The model was trained on data through approximately September 2023. The 2021–2026 period was entirely held out as an out-of-sample test. This includes the full COVID inflation spike — the most severe inflationary episode in 40 years — giving a rigorous stress test of the model's predictive capability.

1-Step-Ahead Backtest — full history and OOS zoom
Figure 1. Rolling 1-step-ahead backtest. Upper panel: full history 1999–2026. Lower panel: OOS zoom 2021–2026. Blue band = 80% prediction interval. Dashed = model median.
In-Sample (pre-2021)

RMSE consistent with published academic benchmarks for monthly CPI forecasting. 80% prediction interval coverage close to nominal — the model is not overconfident.

Out-of-Sample (2021–2026)

RMSE of 3.1pp reflects the 2022 spike — an unprecedented event. Post-peak (2023 onward): RMSE ~0.6pp, comparable to best-in-class econometric models. The model tracked the disinflation path accurately.

Honest caveat on OOS RMSE: The 3.1pp figure is dominated by the March–October 2022 surge that reached 9%. No model trained on pre-2022 data could fully anticipate the magnitude of that shock. The relevant benchmark for forward-looking use is the post-peak RMSE of ~0.6pp, which reflects the model's true steady-state accuracy.

Input Data — 15 Macro Feature Trends

The chart below shows all 15 variables the model ingests, plotted from 1990 through April 2026. The green dotted line marks the Fed's 2% inflation target. The red dotted vertical marks the June 2022 peak. The shaded region is the post-2020 out-of-sample period — data the model never saw during training.

All 15 macro input features — 1990 to Apr 2026
plots/feature_trends.png — all 15 model inputs. Click to zoom.
Price signals (top row)

CPI headline, core, and PCE all show the same 2022 spike and subsequent disinflation. All three are now hovering just above the 2% target — confirming the "slow lane" baseline.

Oil & commodity lags

Oil YoY and its 1-, 3-, and 6-month lags are included separately — because oil shocks take different amounts of time to pass through to consumer prices. The model learns the timing automatically.

Financial & labour signals

Fed Funds, yield spread, TIPS breakeven, UMich expectations, wages, and unemployment give the model visibility into tightening cycles, labour market pressure, and forward-looking inflation expectations.

Understanding Inflation — Measures, Components & Expectations

Not all inflation is the same. The chart below breaks down what inflation is made of, which parts are sticky vs flexible, and whether people's expectations are anchored — three questions that directly determine how long an inflation episode lasts.

The Three Main Inflation Measures
CPI Headline vs CPI Core vs PCE Core — 1999 to Apr 2026. Click to zoom.
What's Driving Headline CPI — component contributions
Approximate component contributions to headline CPI, 2018–present. Click to zoom.
Sticky vs Flexible — Shelter, Services and Goods
Shelter, Services, Core Goods, Medical Care YoY — stickiness comparison. Click to zoom.
Inflation Expectations vs Reality
UMich 1Y inflation expectation vs actual CPI — expectations anchoring. Click to zoom.
① The three measures — what's the difference?

CPI Headline — everything: food, energy, shelter, goods, services. Volatile but complete.

CPI Core — strips out food and energy. Less volatile, better signal of underlying trend. The gap between headline and core tells you how much of inflation is driven by oil and food prices.

PCE Core — the Fed's preferred measure. Uses different weights than CPI (lower shelter weight, higher healthcare weight) and tends to run ~0.3pp below CPI Core. When the Fed says "2% target," they mean PCE.

② What drove the 2022 spike — and what drove it back?

The stacked bar chart shows the approximate contribution of each component to headline CPI since 2018. The 2021–2022 surge was a simultaneous shock across all four components: energy (red), food (gold), goods (purple), and shelter (orange) all spiked together — a once-in-40-year co-movement.

The decline from 9.1% to 2.2% was led by energy going negative, goods deflating, and food normalising. Shelter has been the last holdout — still above 4% as of early 2026 — which explains why core is still above headline's "feel."

③ Sticky vs flexible — why shelter matters most

Core goods deflate quickly when supply chains normalise — we saw this in 2023. Energy is volatile but mean-reverts.

Shelter (rent + owner-equivalent rent) is the stickiest component — it has a 12–18 month lag from actual market rents to CPI measurement. This is why the Fed watches it closely: once shelter inflation gets embedded, it takes 1–2 years to wash out even after market rents stabilise. As of April 2026, shelter is still the primary reason inflation hasn't fully reached 2%.

④ Expectations — the self-fulfilling risk

The University of Michigan 1-year inflation expectation is a leading indicator. When consumers expect prices to rise, they demand higher wages and accept higher prices — making inflation self-fulfilling.

The critical period was 2021–2022: expectations spiked to ~5% but never de-anchored into a persistent wage-price spiral — largely due to Fed credibility. Currently expectations sit at ~3.4% — above actual CPI. If expectations rise further, it becomes a model input for our stagflation scenario.

Why this matters for the model: The InflationTransformer ingests CPI headline, CPI core, and PCE core separately — allowing it to learn the wedge between them over time. Shelter stickiness, energy volatility, and expectation dynamics are all embedded in the training history. When we run the stagflation scenario, the model draws on exactly these historical patterns of component co-movement.

Current Conditions — April 2026

Headline CPI has declined from its 9.1% peak (June 2022) to approximately 2.2% as of April 2026. The disinflationary trajectory has been broadly in line with model projections. Key macro conditions informing the current forecast:

Price Momentum

CPI YoY trending down from 2.7% (Nov 2024) to 2.2% (Apr 2026). The pace of disinflation has slowed — suggesting proximity to a floor near 2–2.5% absent new shocks.

Oil & Commodities

WTI oil flat-to-slightly-declining in 2025–2026. No major commodity supercycle underway. This is the key reason the baseline forecast is benign.

Policy & Trade

Primary upside risk is tariff escalation. Import price pressure has structural lag of 6–9 months to pass through to headline CPI, making M+6 to M+12 the most sensitive window.

Scenario Analysis — 12 Macro Pathways

The scenario engine projects CPI through April 2027 under 12 distinct macro environments. All scenarios share the same starting point (April 2026, ~2.2%) and diverge based on the shocks applied. The structural overlay ensures pass-through is calibrated to historical multipliers rather than relying solely on what the model saw in training.

Tariff / Supply-Shock Scenarios — Origin Apr 2026
Figure 2. Tariff and supply-shock scenario fan chart. White dot = current CPI (Apr 2026, ~2.2%). Dashed lines = median. Inner band = 25–75th percentile. Outer band = 10–90th percentile. Right-edge annotations = M+12 median values.

All 12 Pathways — Small Multiples View

The small multiples chart shows every scenario as its own panel — same y-axis scale, same time axis — so you can compare the shape and speed of each inflationary path at a glance. Scenarios that spike early (oil shocks) look different from scenarios that ramp slowly (tariff pass-through) or fall monotonically (demand slump).

All 12 Scenarios — Small Multiples
Figure 2b. Small multiples — all 12 scenarios on a fixed −2% to 10% y-axis. Grey line = CPI history. Coloured line = median forecast. Bands = 25–75th and 10–90th percentile. Origin: Apr 2026.

Scenario Outcomes — M+12 Medians (April 2027)

Scenario Description M+12 CPI vs. Baseline Risk Tier
Stagflation Oil +30%, PPI +20%, Wages +10%, Import +15% 8.6% +5.9pp EXTREME
Commodity Supercycle Broad commodity surge across energy + metals 6.5% +3.7pp EXTREME
Tariff — Severe Import +25%, PPI +15% (full escalation) 6.2% +3.5pp EXTREME
Tariff — Moderate Import +15%, PPI +8% (partial escalation) 5.0% +2.3pp HIGH
Tariff — Light Import +8%, PPI +4% (limited pass-through) 3.9% +1.2pp HIGH
Oil +20% WTI oil rises 20% YoY over forecast horizon 3.7% +1.0pp ELEVATED
Oil Flat (Baseline) No new shocks — status quo macro environment 2.74% MODERATE
Oil −10% WTI oil falls 10% YoY (easing energy costs) 2.2% −0.6pp MODERATE
Demand Slump Sharp demand contraction, spending & imports fall 0.3% −2.4pp LOW

Probability Distributions at Key Horizons

The distribution snapshot chart shows the full probability range for each scenario at four specific dates — not just the median. Wider bars indicate greater model uncertainty; bars that overlap mean those scenarios are indistinguishable at that horizon.

All Scenarios — Distribution Snapshots at M+3 / M+6 / M+9 / M+12
Figure 3. Horizontal box plots at M+3 (Jul 2026), M+6 (Oct 2026), M+9 (Jan 2027), M+12 (Apr 2027). Bar body = 25–75th percentile. Whiskers = 10–90th percentile. Vertical tick = median.
Near-term (M+3, Jul 2026): Almost all scenarios cluster between 2.4–3.6%. Structural pass-through of any shock is only ~20% complete at 3 months. It is not yet possible to distinguish tariff_light from oil_up_3pct at this horizon.
Medium-term (M+9–M+12): Scenarios diverge dramatically. The spread from lowest (demand_slump, 0.3%) to highest (stagflation, 8.6%) reaches 8.3pp — a portfolio-relevant range that demands scenario-weighted positioning from Q4 2026 onward.

Upside Inflation Risk at M+12 — Probability Analysis

This section answers the question most relevant for portfolio risk management: given a scenario, what is the probability that CPI exceeds a threshold that would force a Fed policy response?

All Scenarios — Upside Inflation Risk at M+12
Figure 4. Probability of CPI exceeding 2%, 3%, 4%, and 5% at M+12. Sorted by P(CPI>3%). Probabilities approximated from 10th/25th/50th/75th/90th percentile outputs.

Risk Probability Summary

Scenario P(CPI>2%) P(CPI>3%) P(CPI>4%) P(CPI>5%) Tier
Stagflation 90%90% 90%90% EXTREME
Commodity Supercycle 90%90% 90%85% EXTREME
Tariff — Severe 90%90% 90%82% EXTREME
Tariff — Moderate 90%90% 78%50% HIGH
Tariff — Light 90%77% 47%25% HIGH
Oil +20% 90%72% 44%21% ELEVATED
Oil Flat (Baseline) 73%43% 21%10% MODERATE
Oil −10% 55%32% 14%10% MODERATE
Demand Slump 15% 10% 10% 10% LOW
Striking finding: Even the no-shock baseline (oil_flat) shows 43% probability of CPI remaining above 3% at M+12. The model is signalling that current momentum alone — without any new policy shock — is sufficient to keep inflation persistently above the Fed's target in nearly half of simulated paths.

Key Findings & Strategic Implications

1
Tariff escalation is the single largest controllable inflation risk.

Every additional "tranche" of tariff escalation adds approximately +1.1pp to the M+12 CPI median. The tariff pass-through channel (import prices → producer prices → consumer prices) has a structural 6–9 month lag, making it invisible in near-term data but highly predictable at the 12-month horizon. Once tariffs are in place, the inflationary effect cannot be reversed quickly — the model shows pass-through continuing to compound through M+12.

2
Stagflation risk is non-linear and back-loaded.

At M+3, the stagflation scenario (3.6%) looks only modestly worse than tariff_severe (3.2%). But by M+12 the gap has grown to 2.4pp (8.6% vs 6.2%). The wage-price feedback loop only compounds after month 6. Positions hedging against stagflation need to be in place before M+6 — not after the first signs appear in the data.

3
Commodity supercycle outperforms tariff-severe at medium horizon.

A broad commodity boom (oil + metals + agricultural) ultimately produces higher inflation than trade tariffs alone — 6.5% vs 6.2% at M+12. Tariff shocks are concentrated in import price channels; commodity supercycles hit energy, food, industrial inputs, and transport simultaneously, creating broader and more persistent pass-through.

4
The path back to 2% requires active demand destruction.

Of all 12 scenarios, only oil_down_10pct (2.2%) and demand_slump (0.3%) project CPI near or below the Fed's 2% target. Passive macro policy will not be sufficient — returning to target requires either a sustained energy price decline or a significant demand contraction. The model assigns near-zero probability to spontaneous disinflation under current conditions.

5
Risk-weighted P(CPI>4%) is already elevated under any scenario mix.

Assume a conservative scenario mix: 50% baseline, 25% tariff_moderate, 15% oil_up_20pct, 10% demand_slump. The probability-weighted P(CPI>4%) = 0.50×21% + 0.25×78% + 0.15×44% + 0.10×10% = 37%. Over one-third probability of a Fed-mandating overshoot before any above-baseline scenario is even assumed likely. At 40% tariff_moderate weight, this rises to ~44%.

6
Short-horizon signals are unreliable for scenario discrimination.

At M+3 (July 2026), the spread across scenarios is only 1.2pp. Investors attempting to read incoming CPI prints through Q2–Q3 2026 as confirmation of a scenario will find the signal too weak to act on. The model suggests positioning based on scenario probabilities (policy signals, trade flow data, wage surveys) rather than waiting for CPI data to confirm — by which time the M+12 outcome is already largely determined.

Illustrative Risk-Weighted Outlook

The table below illustrates a plausible scenario probability assignment and the resulting risk-weighted CPI distribution at M+12. This is not a point forecast — it is a weighted average of conditional distributions based on assumed scenario likelihoods. The CIO should substitute their own probability assignments.

Scenario Scenario
Weight
M+12
Median
Weighted
Contribution
P(CPI>4%)
Contribution
Oil Flat (Baseline) 40% 2.74% 1.10pp 8.4%
Tariff — Moderate 25% 5.00% 1.25pp 19.5%
Oil +20% 15% 3.74% 0.56pp 6.6%
Tariff — Light 10% 3.88% 0.39pp 4.7%
Demand Slump 10% 0.27% 0.03pp 1.0%
Risk-Weighted Total 100% 3.33% 3.33pp 40.2%
Interpretation: Under this illustrative mix, the probability-weighted CPI median at M+12 is 3.3% — 1.1pp above current — with a 40% probability of exceeding 4%. This warrants meaningful inflation protection in portfolios, particularly for instruments sensitive to the M+6 to M+12 window (TIPS, commodity exposure, floating rate credit).

Limitations & Model Boundaries

What the model does well

Cyclical inflation dynamics, pass-through from oil/commodities, tariff and import price channels, disinflation paths following demand normalization. All channels well-represented in training data (1998–2024).

Known limitations

Fiscal dominance (large deficit monetisation), geopolitical supply disruptions (wars, sanctions), structural breaks in wage formation, financial crises with non-linear credit effects. Scenarios outside the training distribution will be under-estimated.

Scenario probability assignment

The model does not assign probabilities to scenarios — that is the user's judgment call. The model only provides P(CPI outcome | scenario). Scenario weights should reflect the team's macro view.

Retraining cadence

Model should be retrained quarterly or when macro regime signals a structural break (e.g., sustained CPI deviation >1pp from model for 3+ consecutive months). Current checkpoint: trained through Sep 2023, with live data inference through Apr 2026.