The Recession Path: How Income, Investment and Credit Can Turn a Technology Boom Into an Economic Reset

Six numbers to keep in view

The forward cycle model: what it actually predicts

Pattern Nexus research note 03. Data frozen October 16, 2026. U.S. projection dates run from inferred 2026 Q3 through 2031 Q3.

The central forecast is continued growth with a meaningful recession risk. The data do not establish that a technology-driven depression is imminent. I now have a runnable, linked model that projects the economy forward. Its baseline is about 2.0% real GDP growth over the following four quarters, with real wage/salary income growing 1.3% and consumption 2.5%.

The next-year recession number has only modest historical forecasting skill. Its calibration-only 95% resampling range is roughly 8% to 24%, and that range excludes several additional sources of model and data uncertainty. It should be read as an approximate risk estimate.

This research completes the dynamic-model build. It preserves weak results and historical misses, including 1928, instead of treating an attractive stress chart as validation. The package contains the original work, model code, coefficients, data and forecast-by-forecast results.

What changed from the earlier research

My earlier research established component relationships, recession classifiers, historical episode comparisons and a broad source archive. They did not turn those parts into one recursively evolving economy. That gap is addressed here.

The U.S. system has 16 endogenous quarterly states. The long historical system has 10 annual states in 18 countries. The annual system is a separate test of long-run prediction, because its wage, output and credit measurements are not identical to the modern U.S. series.

This is a predictive model informed by economic mechanisms. It does not identify structural causal effects. A linked set of regressions can simulate feedback while still missing a structural break, a bank run or a regime that is absent from its training sample.

What was not assumed

No fixed cycle length, predetermined crash date, inevitable technological unemployment, or automatic conversion from an AI investment boom to a depression was imposed. The model must produce those implications from observed relationships if the data support them.

The current GDP projection

The model nowcasts 2026 Q3 GDP growth at 2.6% at an annual rate. This is an estimate, not an observed GDP release. Future growth starts after that nowcast quarter; it does not count the nowcast as an extra forecast quarter.

The baseline never develops an endogenous output collapse over the displayed five years. Its eventual return toward historical average growth is a model feature and a limitation: a lasting change in technology, institutions or financial structure could invalidate that mean reversion.

Spending runs ahead of wage income

Over the next four quarters, deterministic consumption growth is 2.5%, real wage/salary income 1.3%, and payroll employment 0.9%. This combination is compatible with a slowing household income engine while aggregate demand continues to expand.

The gap is not an accounting estimate of new household borrowing. Wage/salary disbursements exclude transfers, proprietor income, investment income and wealth effects. Aggregate consumption and this wage measure cover different income channels. The model captures their historical association; it does not prove that households can fund this divergence indefinitely.

How much do the model families disagree?

The candidate families differ in the number of lags, which cross-variable links are allowed and whether credit, investment and income divergences enter as nonlinear terms. Every family currently projects positive growth. The selected joint four-lag model is toward the lower end of the one-year forecasts.

Agreement on the sign of the baseline does not prove that a recession will be avoided. All these candidates use related observations, share revised historical data, and largely extrapolate within the same broad regime. Their disagreement understates uncertainty about an unprecedented structural change.

The selection rule is retained even though its historical joint activity loss is 2.3% worse than the AR benchmark. Choosing a current model does not certify that the choice will outperform.

How the linked economy evolves

At each quarter, the model predicts every state from earlier values of all permitted states. Future predictions are then inserted into the history and used to predict the next quarter. Spending therefore responds to projected income and finance; investment responds to projected demand and credit; employment and income respond to projected output and productivity.

Standardized system:

z(t) = intercept + A1 z(t-1) + ... + Ap z(t-p) + B f(history) + innovation(t)

The nonlinear candidate recomputes its buildup terms from the simulated history at every step. It does not forecast one static credit signal independently and then attach a recession label to it. The selected current model is linear with four lags; the nonlinear candidates remain visible for comparison.

These are estimated predictive links, not a complete national-accounts identity system. They do not independently enforce government, external-sector and household budget constraints. That distinction matters for causal policy claims and for very long stress scenarios.

What each U.S. state means

Growth variables use 100 times the log change per quarter; they are not annualized. Rates use percent or percentage points. Investment is transformed to keep its simulated share between zero and one. Credit and digital shares use logs. Full series identifiers and transformations are in the data dictionary and code. ^{[1] [2] [3] [4]}

The digital proxy is broad capital expenditure, not an AI adoption measure. It covers a usable modern history starting in 1959, not 150 years of consistent technology measurement. The model cannot infer a century-long technological clock from that proxy alone.

The information clock prevents a false forecast

The historical forecasts use the same relative observation delays. Quarterly future paths begin after the news quarter. A one-year forecast made with 2026 Q3 news therefore ends in 2027 Q3. That timing is applied identically to all candidate models and baselines.

This is a pseudo-real-time experiment on revised data. It is not a claim that every series was available in its final revised value on the historical information date. Complete joint-system vintages were not reconstructed. Previous real-time component tests remain in the archive but do not remove this limitation.

An audit found that a nine-month NBER label delay was too short for some historical announcements. The final diagnostic uses two years and filters the unannounced state. NBER reports delays as long as 21 months since the formal committee process began. ^[5]

Two isolated shutdown-month gaps, October 2025 CPI and unemployment, were interpolated for completed-quarter aggregates and are explicitly flagged. Missing calendar periods were not compressed out of the time series.

What was fitted and how it was chosen

Every candidate is fit with shrinkage values 1, 10, 100 and 1,000 on standardized predictors. Coefficients are shrunk toward own-state persistence: 0.8 for level-like financial/share states and 0.2 for growth states. A two-pass robust weight reduces the influence of extreme fitting rows; raw residual tails are retained for simulation.

Eight nonlinear terms represent five-year credit buildup, investment-share buildup, productivity-pay divergence, spending-wage divergence, rate-reset exposure, investment-credit interaction, divergence-credit interaction and an inflation-adjusted restraint proxy. Training-scale hyperbolic tangents bound these features; forecast output is not clipped.

Each family's shrinkage, and the overall family choice, use only previous four- and eight-quarter activity errors whose eight-quarter windows are complete. The target loss combines GDP, consumption, wages and payrolls after past-only scale normalization. There must be 16 mature earlier origins before selection departs from the initial rule.

The architecture was designed after earlier research had already exposed historical outcomes. Chronological algorithmic evaluation is valuable, but this is not an untouched human research holdout or a preregistered discovery claim.

GDP forecast accuracy: gains and limits

The joint four-lag family reduces one-year GDP RMSE from 2.63 to 2.28 log percentage points, about 13%. Its five-year GDP RMSE is worse than AR. The past-selected rule is slightly worse than AR at one year. A fixed equal-weight committee improves the point estimate by about 11% at one and two years, but its block uncertainty includes no gain.

The joint forecast claim must survive uncertainty

The fixed joint four-lag family shows a 22.4% reduction in normalized joint RMSE, while the past-selected family rule worsens it by 2.3%. The distinction is important: inspecting which fixed family performed best after the fact is not the same as knowing which family to choose before a forecast.

Every interval shown includes zero improvement. Removing the pandemic forecast windows weakens the apparent gains. The model is useful as an explicit conditional forecasting system, but the evidence does not support a broad claim that its selected baseline reliably dominates a simple AR forecast.

Forecast windows overlap. I resample paired calendar blocks of 20 quarters so that a five-year episode is not counted as many independent discoveries. Alternative eras and exclusions remain in the results files. A large country-quarter or origin count does not create many independent crises.

Intervals are not automatically calibrated

For the next year, the selected raw GDP simulation has an 80% interval of roughly -0.3% to +4.4%. Using the past-only marginal error scale widens it to -0.7% to 4.8%. This widening does not turn a weak model into a strong forecaster.

At five years the selected raw nominal 95% interval covered only about 87% of historical outcomes. The simple past-error adjustment still did not restore 95% coverage. The current error-scaled five-year 95% GDP interval is approximately 0.9% to 25.2%; it remains conditional on the model and the historical regime.

Simulation uncertainty includes empirical joint innovation blocks and conditional parameter resampling. It omits uncertainty about future policy rules, financial architecture, technology measurement and a genuine regime break. Marginal scaling does not supply coherent calibrated probabilities for every whole-path event.

What happened around actual downturns?

The model picks up some slowdowns and conditions on rapidly changing news, but it does not reliably forecast every recession in advance. Rebounds after a collapse are a different task from predicting the collapse before it begins. The pandemic contribution to the model's average accuracy must therefore be interpreted carefully.

The archive keeps the GFC, 2001, 1990-91, 1981-82 and pandemic forecast rows. It also keeps non-recession periods when the model was too pessimistic. Choosing a few visually successful episodes would be a weaker test than the full rolling record.

A one-year growth forecast can be positive even if an NBER recession occurs during part of the year. Conversely, two negative GDP quarters need not coincide exactly with an NBER recession. I evaluate these targets separately because they measure different economic outcomes.

How economic paths become a recession diagnostic

A separate logistic transition/emission map uses projected GDP, consumption, wage growth, payroll growth, the unemployment change and the lagged recession state. It is trained only on labels at least two years old. Each simulated activity path then generates a recession-state path. The mapping is not the source of the economic dynamics.

The raw mapping is badly upward-biased: the selected system averages about 36% predicted one-year onset risk against an observed rate near 12% across expansion origins. A one-intercept odds correction, trained only on earlier completed and recognized event windows, reduces that bias.

This recalibration is an exploratory repair made after inspecting the raw failure. The full sequential results are reported. The initial shorter-label-lag diagnostic is retained in an audit folder and is not used for final probability claims. ^[5]

The recession estimate and its historical support

The frequency benchmark uses all earlier recognized NBER expansion-origin windows, not a 50% starting guess. The selected diagnostic improves the Brier score by about 6% in this experiment. Paired block uncertainty includes deterioration, so this is modest evidence rather than a robust forecasting victory.

An estimate around 15% supports meaningful risk monitoring. It does not make recession the base case, and it does not imply a 15% chance of depression. Longer-horizon onset probabilities naturally rise because more quarters allow more opportunities for an ordinary recession.

The 150-year test is a separate economy

The annual panel covers 1870-2020. Rolling forecasts begin in 1900 and end in 2019, with horizon-specific maturity limits. It uses output, consumption, a real wage index, a GDP-per-capita productivity proxy, investment, credit, inflation, rates, the curve and real money growth. Country means are partially pooled with 25 global pseudo-observations. ^[6]

The primary fit and score exclude the two world-war windows and histories contaminated by them. A separately saved all-calendar sensitivity retains war outcomes. The Great Depression is retained. Parameter resampling uses calendar-year blocks common across countries; innovation blocks remain within a country's consecutive history.

The selected annual system does not beat AR on the joint real-activity target: its joint RMSE is about 1.3% worse. The long record expands the range of regimes tested, but it does not rescue an unsupported technological clock.

The 1928 forecast is a necessary falsification

From the 1928 information set, the selected annual system projects about +15.2 log percentage points of output growth over five years, approximately +16.4% in ordinary percentage terms. The realized five-year outcome is -27.6 log points, approximately -24.2%. The raw interval misses the collapse decisively.

The useful inference is narrower: observed income, investment, credit and policy states allow explicit conditional paths, but the model lacks reliable advance identification of a systemic break of this scale. The historical distinction between an ordinary downturn and a depression must remain visible in the conclusion.

Can the model quantify depression risk?

The U.S. quarterly path target instead requires a 10% loss and the final eight quarterly endpoints below the starting level. Different persistence definitions and populations are not pooled into one probability. The quarterly post-1959 record contains no comparable sustained 10% loss from these forecast origins.

Does the system contain a technological clock?

A complex eigenvalue can produce an oscillatory response even in a purely predictive model. Its angle defines a mathematical period; its magnitude defines how rapidly the response fades. Finding a period is not evidence that the economy is marching toward a scheduled crash.

The longest prominent current oscillatory mode has a period near 15 years but an amplitude half-life around one year. A disturbance fades substantially before completing that oscillation. The model therefore does not supply a durable 15-year reset clock, much less a validated century-scale technological timetable.

Technology adoption and financing structures have changed across electrification, mass production, computing and AI. The comparable aggregate technology measurement is much shorter than the full macrohistory panel. Too few independent technological transitions exist to estimate precise calendar probabilities for a massive reset from this dataset.

Does the technology channel improve forecasts?

Removing the digital investment share barely changes GDP forecasting performance. Its one-year point improvement is roughly 0.6%, with an interval that includes no gain. Financing and labor-channel removals also have intervals spanning zero; some point comparisons favor the simpler ablated system.

This does not prove technology, employment or finance are economically irrelevant. It says their incremental forecasting contribution is not clearly isolated in this highly correlated, strongly regularized aggregate system. The model is not evidence that a productivity-pay divergence mechanically causes a near-term collapse.

The channel tests and fixed committee were added as exploratory robustness checks after the original tournament was inspected. Their past-only fitting is retained, while their research-stage status is disclosed. They are not new untouched holdouts.

What puts the baseline into recession territory?

The main result is the combination. Income weakness alone reduces growth; an investment reversal brings it close to zero; financing pressure adds risk. Together, the stated conditions produce a contraction of roughly 1% over the first year. Their joint effect is an associated conditional path, not a sum of separately identified causal effects.

The exact stress assumptions

The model conditions joint innovations by their estimated covariance. This allows observed or imposed values to inform other current states. It does not identify a structural shock or prove what a policy intervention would cause.

A lower credit/GDP ratio is not synonymous with a lower nominal credit stock. BIS divides end-quarter credit by the sum of four quarters of GDP. Both numerator and denominator can move. The financing case is therefore a joint financial-condition experiment, not a quantified bank lending withdrawal. ^[4]

The easing variant does not materially prevent the first-year contraction and slightly worsens it in the conditional model. That result reflects reduced-form associations, including policy easing during weak activity; it is not evidence that easing causes recession. A causal policy comparison would require separately identified shocks.

A concrete conditional break-even

The combined break-even is about 0.68 times the specified package. This corresponds to an initial quarterly wage-growth shortfall around 0.34 log point, a private investment/GDP share about 1.36 points lower, a corporate spread about 0.68 point higher and a credit/GDP ratio about 6.8 points lower, with the stated timing.

Those numbers make the bearish thesis testable. They are not observed facts, independently identified causal thresholds or estimates of how likely the package is to occur. The surface is conditional on all remaining modeled relationships and the selected shrinkage.

From recession stress to depression-like persistence

The 3x case imposes an investment/GDP share 6 percentage points below baseline, a corporate spread 3 points higher, credit/GDP 30 points lower, and wage growth 1.5 log points lower for the first four quarters, then 0.75 lower. It also imposes the corresponding much weaker digital investment path. These are large, sustained assumptions.

The model produces a roughly 12% five-year output loss under those conditions. That establishes a conditional route to depression-like output persistence. It does not estimate the probability that the conditions will arise, and the scenario lies beyond well-validated tail forecasting.

Why duration changes the economic outcome

A sharp but temporary downturn can be followed by a recovery. A depression-like result requires weakness to keep propagating through income, capital formation and finance. The duration grid makes that distinction explicit: the same package can produce a small first-year contraction and then recovery, or a much larger cumulative loss if stronger conditions persist.

The unconditional probability of a persistent loss would require both the chance of a structural stress regime and the chance of persistent loss within that regime. This exercise estimates conditional paths, not the regime's future occurrence rate. Assigning it an arbitrary scenario weight would create a number without establishing the requested historical prediction.

The long-history test does not provide reliable timing or ranking of the strict persistent-loss event. The correct empirical conclusion is uncertainty about depression odds, alongside a quantified mechanism for what would be required to produce one.

Where the liquidity and gold work fits

The PN liquidity framework distinguishes household purchasing power from collateral and balance-sheet capacity. The integrated model retains real money growth, credit, the curve, spreads and rates as endogenous financial states. It does not pretend that WALCL, TGA, reverse repo and stablecoin plumbing form a consistent 150-year dataset. ^[7]

The four-reservoir LCI starts in a much shorter institutional regime. Its relationship to gold or other assets is not automatically a recession forecast. The retained tests show weaker performance at six and twelve months, with longer-horizon improvement based on very few independent windows. Asset repricing and falling real household demand can coexist.

Gold also crosses monetary regimes in the long historical record, including periods with an official price peg. A smooth century-long price series does not supply a comparable freely traded gold-return signal throughout. CPI, gold and liquidity are therefore distinct measured channels; none is used as a stand-alone proof of an inevitable technological reset.

For the framework behind the liquidity distinction, see my Hard Assets Follow Liquidity: LCI Framework White Paper. The linked economy here gives those financial states an explicit place alongside household demand and investment. ^[7]

Feedback responses and the limit of causal language

Some responses fade rapidly; some have counterintuitive signs. In the selected model, imposing higher productivity and a higher digital share does not automatically raise employment or output. Productivity can coincide with weak labor demand, and capital expenditure can be responding to other conditions.

A convincing causal account would need identification of technology, credit supply and policy shocks, distributional spending behavior and balance-sheet losses. Aggregate predictive feedback is a useful step, but it does not resolve those identification problems. That is why I describe these results as conditional paths.

What would change the conclusion?

The present conclusion would become more bearish if the observed data began matching a persistent combination of weak aggregate real wage income, broad employment deterioration, investment retrenchment and financing stress. A technology-spending boom by itself is insufficient in these tests.

For a prospective evaluation, freeze this model version and its current forecast paths now. When new outcomes arrive, score their growth errors, interval coverage and recession calibration before redesigning the system. A future evaluation of the frozen model would be stronger than another retrospective search for a pleasing cycle.

No automatic monitoring or publishing has been set up. The delivered CLI reruns the frozen data; updating to a new information date requires refreshing source observations and auditing their release timing.

Where I stand: the recession case becomes stronger when weakening real income, employment, capital spending and finance begin reinforcing one another. This model gives me explicit paths and thresholds to compare with incoming evidence. Its current central path remains expansion, and its historical record does not establish that a depression is approaching. That is the conclusion I will carry forward and test against the next observations.

Reproduce and inspect the forward model

The full package retains the earlier research alongside phase3. The new model has no external API dependency. Its essential numerical dependencies are NumPy, pandas, SciPy and scikit-learn. The frozen coefficients and input histories are saved so that the projection can be reviewed independently of this article.

python phase3/scripts/run_model.py --information-quarter 2026Q3 --out-dir my_forecast

Checks verify that changing future economic observations does not change an earlier forecast, changing immature target outcomes does not change selection, and changing unannounced NBER labels does not change the mapping. The saved current coefficients reproduce the current projection. Workbook results are checked against Python and a wage innovation is verified to propagate into GDP.

Technical score and uncertainty definitions

Linear companion matrices are checked and, if necessary, shrunk to a spectral radius no greater than 0.995, with all adjustments recorded. Nonlinear terms are bounded; forecast output is not clipped to conceal explosive paths. A stable linear matrix is not a guarantee of economically correct long-run dynamics.

Questions readers may have

Is recession the model’s base case?

No. The baseline is about 2.0% real GDP growth over the four quarters after the inferred 2026 Q3 starting point. The separately evaluated recession diagnostic is about 15.5%, conditional on the starting quarter being an expansion. Positive growth over a full year can coexist with a recession during part of that year.

Can this research put a reliable probability on a depression?

The historical validation does not support a precise current depression probability. The strict annual persistent-loss test has only five positive country-origins across three origin years, poor ranking performance and a large miss from the 1928 information set. I retain those failures alongside the scenario results.

What combination turns the baseline into a contraction?

The declared income, investment and financing squeeze produces about a 1% first-year GDP contraction. About 0.68 times that combined package takes deterministic first-year growth to zero. The package specifies the timing of wage, investment, digital-share, spread, short-rate and credit conditions; it is not an estimated probability or an independently identified causal threshold.

Does the model prove a technological reset cycle?

No. The broad digital-investment proxy is not a direct measure of AI adoption, its incremental forecasting contribution is small and uncertain, and the roughly 15-year mathematical oscillation loses half its amplitude in about one year. A persistent calendar cycle is not established.

How do gold, CPI and the liquidity framework fit?

CPI helps measure real purchasing power, while money, credit, rates, spreads and the curve enter the linked economy. The LCI and gold tests supply related asset-market context. Short modern plumbing histories and changes in the gold monetary regime prevent those series from becoming one consistent 150-year recession signal.

What should readers watch next?

The combination matters: aggregate real wage and salary income, employment breadth, consumption, private and digital investment, corporate spreads, credit quantities and debt service. A persistent deterioration across these channels would strengthen the recession thesis. Future results should be scored against this frozen model before its design is changed.

Sources and evidence provenance

All figures and model estimates in this article are the report’s phase3 calculations unless explicitly labeled as retained phase2 work. Source snapshots, retrieval records and hashes from the earlier phases remain in the package. The phase3 manifest hashes the new inputs, code and outputs. The research data cutoff remains October 16, 2026.

1. Federal Reserve Bank of St. Louis, FRED
   Distribution of the U.S. BEA, BLS, Federal Reserve and BIS series; the data dictionary records individual identifiers.

2. Bureau of Economic Analysis, national accounts
   GDP, private investment, personal consumption and wage/salary disbursements.

3. Bureau of Labor Statistics
   Employment, unemployment, CPI, productivity and compensation.

4. BIS total credit documentation
   Broad private credit includes nonbank and cross-border financing; credit/GDP uses end-quarter debt and four-quarter GDP.

5. NBER business-cycle dating FAQ and announcements
   The formal dating committee began in 1978; documented announcement delays range from four to 21 months.

6. Jorda-Schularick-Taylor Macrohistory Database, release 6
   18 advanced economies since 1870. The model uses the frozen panel through 2020; data definitions and licensing are retained with sources.

7. Pattern Nexus liquidity framework
   My research on balance-sheet reservoirs, collateral liquidity and hard-asset repricing; the earlier thematic inventory retains the wider corpus.

8. Afanasyeva (2020), Federal Reserve FEDS 2020-045
   Methodological context: multivariate credit forecasts and financial-crisis signals. Its published validation is not validation of this model.

9. NBER business-cycle dates
   Reference chronology. The report separately evaluates GDP contraction and persistence targets.

Back to top