AI / INVESTMENT INTELLIGENCE
SOXX, looking three months ahead.
A testable return forecast, transparent validation and the industry evidence behind the investment view.
MINGYANG FORECAST LAB / V2.0Monthly model · 3-month horizon
RESEARCH SIGNAL / LIMITED EVIDENCEA forward view.
A forward view.
A visible test.
Predict SOXX’s next-quarter return from market conditions. Test whether industry changes add information.
SELECTED MODEL Market signals
Information date 2026-08-31 · constructed 2026-09-13
Reconstructed month-end forecast; it was not published on August 31.
MODEL FORECAST / NEXT 3 MONTHS+8.2%
SOXX dividend-adjusted return
31 Aug → 30 Nov 2026
Historical residual range-7.7% to +26.5%10th–90th residual percentiles; not a guaranteed coverage interval.
HOW MUCH CONFIDENCE?
Positive historical improvement. Low conviction.
The selected model beat the earlier version and the historical mean, but its uncertainty interval includes no improvement. It called the direction correctly in 78.6% of test windows—the same rate as always predicting a gain.
Range calibration needs work
The nominal 80% residual band contained only 50% of later outcomes. It is displayed as a historical range, not an 80% confidence promise.
3 of 4 current market inputs are more than two training standard deviations from their means. Today’s setup is unusual.
EXPLAIN THE NUMBER
Contributions in percentage pointsWhat moves our forecast?
Historical-return anchor+5.3 pp
Next-quarter return forecast+8.2%
The contributions sum to the forecast. They describe this fitted model, not causal effects on stock returns.
LOCKED LATER-PERIOD CHECK / 2023–2026
Forecasts meet outcomes.
Exact forecast and outcome observations
| Forecast origin | Outcome date | V2 forecast | Historical mean | SOXX outcome |
|---|---|---|---|---|
| 2023-01-31 | 2023-04-28 | +4.7% | +4.0% | +2.4% |
| 2023-04-28 | 2023-07-31 | +4.8% | +4.1% | +30.1% |
| 2023-07-31 | 2023-10-31 | +5.1% | +4.3% | -17.1% |
| 2023-10-31 | 2024-01-31 | +3.8% | +4.1% | +32.7% |
| 2024-01-31 | 2024-04-30 | +4.1% | +4.3% | +9.7% |
| 2024-04-30 | 2024-07-31 | +5.1% | +4.5% | +10.1% |
| 2024-07-31 | 2024-10-31 | +4.8% | +4.5% | -7.0% |
| 2024-10-31 | 2025-01-31 | +3.9% | +4.4% | +0.1% |
| 2025-01-31 | 2025-04-30 | +3.6% | +4.3% | -15.6% |
| 2025-04-30 | 2025-07-31 | +4.6% | +4.2% | +30.9% |
| 2025-07-31 | 2025-10-31 | +4.4% | +4.3% | +28.0% |
| 2025-10-31 | 2026-01-30 | +6.1% | +4.5% | +13.1% |
| 2026-01-30 | 2026-04-30 | +6.0% | +4.6% | +33.3% |
| 2026-04-30 | 2026-07-31 | +6.9% | +4.8% | +9.5% |
14 non-overlapping 3-month return windows. Selection used earlier dates; the later period was already visible in V1 research. This is a retrospective experiment, not a pristine holdout or a live track record.
SHOW EVERY CANDIDATE
Did the industry data improve the prediction?
Not in this test.
The market-only model won earlier selection. Adding industry changes made the later-period squared error 0.9% worse than market-only. Industry indicators remain an interpretation and thesis-monitoring layer.
| Fixed candidate | Earlier validation · error improvement | Later period · error improvement | Role |
|---|---|---|---|
| Historical mean | 0.00% | 0.00% | Common benchmark |
| Market signals | +0.61% | +3.12% | Selected before later test |
| Industry changes | -0.75% | -0.96% | Not selected |
| Market + industry | +0.31% | +2.25% | Not selected |
Positive percentages mean lower squared error relative to the same historical-mean benchmark. Negative means worse. All candidates use common dates; there was no search for the best later-period result.
Fit and compare
Minimum 60 matured observations. Expanding fits, four market features and four industry changes. Fixed regularization.
Select once
January 2013–July 2022 selection origins. The July forecast resolves in October; the next disjoint origin would end after the cutoff.
Evaluate without retuning
January 2023 onward. Compare errors, uncertainty, direction and range coverage. Keep failed tests visible.
Start the real test
Research inception 2026-09-13. Zero completed prospective forecasts. Future vintages must be logged before outcomes.
Forecast specification, empirical probabilities and source data
The point forecast blends a fixed-penalty ridge regression and the expanding historical mean, 50/50. Training uses only outcomes known at each origin; feature scaling is also fitted on training data only. Macro observations are delayed 60 calendar days.
Unrounded empirical up estimate: 78.9%, based on 225 prior matured prequential residuals. This is not a calibrated confidence rating. Later-period Brier score 0.171 versus 0.181 for the historical frequency; only 14 evaluation windows.
The separate next-63-day volatility model has -2.6% error skill versus its expanding-mean baseline. It does not establish a risk-forecast edge and is not substituted for return prediction.
- Economic inputs use revised series and a fixed60-day lag, not historical release vintages. STLFSI4 itself was reconstructed under an updated methodology.
- SOXX has changed benchmark methodology during this ETF history; results describe the fund, not a fixed semiconductor constituent basket.
- Adjusted market prices reflect current vendor revisions; all results use matched dates. No trading execution, taxes or transaction costs are modeled.
- Only a small fixed candidate set was tried. The evaluation period was already visible in prior research. A future live record is required before strong reliability claims.
- No point-in-time earnings revisions or valuation histories were available. They were excluded instead of backfilled.
- The latest completed-month forecast is reconstructed on September13 from August31 information; it was not issued on August31.
- Probability and range are empirical research estimates. Calibration is measured separately; an80% band can miss large outcomes.
Financial Modeling Prep · SOXX and SPY dividend-adjusted daily prices ↗U.S. Census · computers and electronic product new orders ↗Federal Reserve · semiconductor and electronic component output ↗St. Louis Fed · Financial Stress Index ↗iShares SOXX fund definition and benchmark history ↗Campbell & Thompson: forecast restrictions and historical-mean benchmark ↗Goyal & Welch: out-of-sample equity premium prediction ↗
Original AI Cycle Index · V1 conditions gauge and test
MINGYANG AI CYCLE INDEX / VERSION 1.0Measure conditions.
Measure conditions.
Test the signal.
Four long-history public proxies. Fixed equal weights.
A conditions gauge for the AI infrastructure cycle.
Reading 2026-08-31 · source month 2026-06
U.S. proxies include non-AI activity; this is separate from the 16-indicator monitor below.
80.6/ 100
Above trailing norms
+14.6 points vs prior month
25.0 index pointsU.S. Census Bureau · computers and electronic products new orders (A34SNO) ↗
22.9 index pointsFederal Reserve Board · semiconductor and other electronic component production (IPG3344S) ↗
9.7 index pointsU.S. Census Bureau · private data-center construction ↗
22.9 index pointsFederal Reserve Bank of St. Louis · Financial Stress Index (STLFSI4) ↗
THE RESULT / PRIMARY 3-MONTH TEST
Predictive edge not established
The score describes industry conditions. This first test does not demonstrate that it leads SOXX returns; its later-period forecast is less accurate than a simple historical-average forecast.
-0.15 correlation-0.40 to +0.14 95% interval-11.9% later-period forecast skill
SAME DATES / DIFFERENT UNITS
2018–2026 · reconstructed historyIndustry conditions and SOXX
Hover or focus a monthly observation. The top panel is a 0–100 conditions score. The bottom panel is SOXX dividend-adjusted growth of 100. Visual co-movement is not a prediction test.
FORWARD RETURNS / DISJOINT WINDOWS
Does today’s score help predict returns?
Skill = 1 − model forecast MSE / historical-mean forecast MSE. A negative result means the model is less accurate. The later-period check starts in 2023; all rules were chosen in 2026, so this is not a prospective track record.
Methodology, vintage limits and reproducible data
score. For each component, calculate its percentile against the previous 36 monthly transformed observations (current month excluded). Average four percentile scores using fixed 25% weights. A score of 50 is a relative historical midpoint, not a 50% probability of gains.
transforms. Year-over-year percentage changes for orders, production and construction; negative monthly-average financial stress for funding. Higher transformed values raise the score.
timing. Every source month is delayed 60 calendar days from month-end. At each month-end, select the latest common source month eligible by the actual last trading day used for SOXX pricing. A weekend release after that market close cannot enter the score. Do not interpolate missing observations or alter constituent weights.
comparison. SOXX uses dividend/split-adjusted month-end close ratios. The index score is a conditions gauge, not portfolio performance. A normalized overlay is descriptive only; it is not a prediction test.
forwardTest. Primary horizon: 3 months. Also show 1 and 6 months without selecting the best result. Non-overlapping windows are anchored to January (January/April/July/October for 3 months).
forecastTest. Fit forward return = intercept + slope × index on monthly observations whose forward outcome ends by December 2022. Test once on disjoint later-period windows from January 2023. Benchmark is the training-period mean return; skill = 100 × (1 − model MSE / benchmark MSE). Positive skill means lower squared forecast error.
uncertainty. 95% circular moving-block bootstrap intervals, 3,000 resamples with fixed seed; blocks cover at least six months. Intervals describe this reconstructed sample and are imprecise for long horizons.
- Economic series are the latest retrieved revisions, not historical release vintages. A 60-day lag reduces publication timing lookahead but cannot remove revision lookahead.
- Census first published the separate data-center series in 2024 and backfilled it to 2014. Pre-2024 results are not a historically tradable backtest.
- All rules were selected in September 2026. The chronological split is a retrospective later-period check, not a genuine prospective out-of-sample record.
- Nominal construction and orders are affected by inflation. U.S. public proxies include non-AI activity and omit foreign supply chains, AI software revenue, issuer margins and valuation.
- A high industry score can coexist with expensive stocks, supply overshoot or disappointment versus expectations. Market co-movement does not identify causality.
- STLFSI4 uses market-based inputs; the composite should not be described as a purely non-market leading indicator.
U.S. Census Bureau · computers and electronic products new orders (A34SNO) ↗ · Federal Reserve Board · semiconductor and other electronic component production (IPG3344S) ↗ · U.S. Census Bureau · private data-center construction ↗ · Federal Reserve Bank of St. Louis · Financial Stress Index (STLFSI4) ↗ · Financial Modeling Prep · SOXX dividend-adjusted daily close ↗
Full index, observations and test results ↓Prospective record begins 2026-09-13; completed prospective forecasts: 0. The current test does not establish predictive value.