A Prediction Is Not a Forecast
Every existing method for forecasting a univariate time series, from exponential smoothing to the most advanced neural network architectures, performs the same structural operation. The method takes as input a sequence of observations indexed by time and produces as output a continuation of that sequence. The operation appears to use time as an input. It does not. What the operation retains is sequence: the ordinal position of each observation within the series, with no preserved record of the temporal structure that distinguishes one position from another beyond ordinal order.
This is the orientation/measurement collapse expressed as a structural property of a forecasting method. The calendar parameter t enters the input data and is stripped by the operation, which preserves sequential order but discards the measurement information the parameter carried. The output is a continuation of the sequence with no use of, or recovery of, the temporal structure the original parameter encoded. What the method presents as a time-based forecast is a sequence-based forecast with no time in it.
All such operations are Univariate Unidimensional (UU) Operations. They produce Predictions: atemporal absolutes that carry a Dimension 4 value but no Dimension 5 probability structure and no Dimension 6 temporal coordinate. No improvement in the algorithmic sophistication of a UU Operation can restore the measurement function the operation has structurally stripped. The ceiling is not computational; it is dimensional.
Temporal Structural Forecasting (TSF) recovers the measurement function by operating on two structurally independent timelines simultaneously. The sequential timeline supplies orientation. The Dimension 6 seasonal timeline supplies measurement: the structural temporal organization of phenomena independent of sequential position. Orthogonal Projection combines the two UU Operations, one along each timeline, to produce a {4,5,6} Forecast object: a Dimension 4 value, a Dimension 5 Calibrated Probability Band, and a Dimension 6 temporal coordinate. The Forecast is a temporal absolute, not an atemporal absolute. It carries both the value and the temporal position the Prediction structurally cannot.
The empirical program tests this architecture across equity timing and demand forecasting. Historical Results are genuine past forecast performance generated without knowledge of outcomes, structurally distinct from backtests that measure model fitting on known data. The results reported in the studies below were produced by the forecasting architecture, not by parameter optimization against historical data.
TSF White Paper
The Timing Problem: Evidence of Exploitable Temporal Structure in Equity Returns
Results from a 10-Year, 346-Stock Preregistered Validation Study
Half of all tactical allocation funds ever launched are dead. The survivors compound at approximately 5% over 20 years, worse than a static 60/40 portfolio that makes no timing decisions at all. The timing problem is the single largest source of capital destruction in systematic investing, and its source is structural: every existing application of time series forecasting generates a Prediction, not a Forecast. A Prediction is the output of a Univariate Unidimensional (UU) Operation along the sequential timeline, an atemporal absolute carrying a Dimension 4 value but no probability structure and no temporal coordinate. Every method from ARIMA to deep neural networks performs a UU Operation. The category is exhaustive.
Temporal Structural Forecasting (TSF) solves the timing problem by generating a {4,5,6} Forecast: a Dimension 4 Forecast Value, a Dimension 5 Calibrated Probability Band (CPB), and a Dimension 6 calendar coordinate. TSF constructs the Forecast through Orthogonal Projection, combining a UU Operation along the sequential timeline with a second UU Operation along the Dimension 6 seasonal timeline drawn from a proprietary library of 81 seasonal models. The CPB at every position constitutes a Map of Normal. Results are Historical Results, not backtests: each Forecast Value was generated from data preceding its position, without knowledge of the outcome, and a Temporal Firewall enforces this structurally.
The construction was applied to 346 S&P 500 constituents across all 11 GICS sectors over 2016 through 2025 under a preregistered methodology. The full parameter space of 1,386 permutations was tested with none discarded. The single worst performing permutation achieves a 67.7% win rate with a p value of 3.33 × 10⁻¹⁷. Of the 1,386 permutations, 88.0% achieve win rates at or above 75%; 90.6% produce p values below 10⁻³⁰. Filled positions reach the 5% profit target 80% of the time. The top portfolios deliver 18 to 21% CAGR with win rates above 88% and Sharpe ratios up to 0.87. The Structural Integrity Analysis subjected the Forecast to 14 major market disruptions spanning 2007 to 2024; coverage rates recovered after every event without adjustment of any calculation or modification of any parameter. Every result presented in this paper can be independently reproduced.
Download PDF
TSF Methodology
The Dimensional Structure of Forecasting
The Model of Temporal Inertia and a Microscope for Time
The Temporal Firewall
Preregistered Methodologies
Supporting Research
Fama and French (1996) called momentum "the main embarrassment" of their three-factor model—the one timing anomaly their framework could not explain away. We demonstrate that this embarrassment was not an exception but a warning: the entire factor-based methodology for dismissing timing anomalies is invalid. Using 6,560 controlled experiments where paired portfolios hold identical stocks with identical weights entered on identical dates—differing only in exit timing—we show that factor loadings fail equivalence tests at an 86% rate. The failures are not marginal: loading differences exceed 0.45 in magnitude. Random entry strategies with zero predictive content produce massive loading shifts. A 10-stock single-sector control eliminates diversification explanations and still shows 96% failure. These results prove that factor loadings measure return-path covariances, not holdings-based risk exposure. The standard interpretation—that loadings reveal what risks a portfolio bears—is empirically false when timing varies. Every paper that used Fama-French regressions to conclude "alpha explained by factor exposure" for a timing strategy was applying a tool incapable of making that determination. Momentum survived as an acknowledged anomaly only because the broken dismissal tool happened to fail visibly for that case. The Halloween effect, January effect, and hundreds of other timing signals were dismissed by the same invalid procedure. Thirty years of factor-based anomaly evaluation must be reconsidered.
Preregistered: https://doi.org/10.5281/zenodo.18304121
This study tests whether temporal structure exists in equity price data and can be systematically exploited for position entry timing. Using a preregistered methodology, we generate specific limit order entry prices one week in advance for 346 S&P 500 constituents across all 11 GICS sectors. We test 1,386 parameter permutations over the 10-year period 2016–2025, encompassing 6 factor strategies, 7 forecast models, and 3 confidence thresholds.
Results demonstrate statistically significant directional accuracy across all permutations. Mean win rate across the full sample is 80.3% (median 80.9%), with sector means ranging from 73.4% (Consumer Staples) to 86.3% (Information Technology). The maximum p-value observed across all 1,386 tests is 1.91 × 10⁻¹⁶; the minimum is 2.69 × 10⁻¹⁸⁴. No permutation fails to reject the null hypothesis of random directional accuracy at any conventional significance level.
Applied to sector-based position trading portfolios over the out-of-sample period 2021–2025, the methodology generates compound annual growth rates of 18–21% with win rates exceeding 88% and Sharpe ratios of 0.58–0.87. These results constitute evidence against the semi-strong form of the Efficient Market Hypothesis and demonstrate the existence of exploitable temporal patterns in equity returns.
This study subjects the Temporal Structural Forecasting (TSF) Calibrated Probability Band framework to a structural integrity test across 14 major market disruptions spanning 2007 to 2024, from the Lehman Brothers bankruptcy through COVID-19 to the 2024 Yen Carry Trade unwind. For each event, we evaluate whether TSF's 90% Calibrated Probability Bands maintain their calibration by comparing event month out-of-bounds rates against individualized, stock-specific baselines derived from each security's pre-shock behavior across the full S&P 500 universe.
TSF maintains structural integrity (pass rate ≥80%) at the event month for 7 of 14 events, including Bear Stearns (90.7%), the March 2009 generational market bottom (89.7%), the European Debt Crisis (90.3%), Brexit (90.3%), Russia-Ukraine (84.3%), SVB (86.2%), and the Yen Carry Trade unwind (86.7%). Excluding the three most severe disruptions (Lehman, COVID-19, U.S. Debt Downgrade), the average event month pass rate across the remaining 11 events is 76.3%. A near-universal recovery signature is observed: 13 of 14 events exhibit monotonic improvement from the point of maximum disruption, with coverage rates returning to pre-shock levels without adjustment of any calculation or modification of any parameter.
The March 2009 result provides the decisive proof point: six months after near-total structural failure during the Lehman crisis (7.3% pass rate at M+1), TSF achieved 89.7% structural integrity at the generational market bottom and returned to 99.0% the following month. These findings establish that the temporal structures identified by TSF represent persistent features of equity market behavior that recover systematically from even the most extreme exogenous shocks.
Conventional portfolio theory holds that diversification improves risk-adjusted returns by reducing idiosyncratic risk. This paper examines whether TSF timing signals alter this relationship. Using low volatility portfolios ranging from 10 to 100 positions across two universes (346-stock Broad and 168-stock Defensive), we find that native (untimed) strategies confirm the diversification benefit: N100 outperforms N60 by 0.6 percentage points CAGR over the full 20-year sample. However, TSF-timed strategies show the opposite pattern: N60 matches or exceeds N100 returns in most periods, and more concentrated portfolios (N40) often deliver the highest CAGR. Over 2006-2025, Broad Universe TSF N40 delivered 7.9% CAGR versus 7.4% for TSF N100—concentration outperformed diversification by 0.5 percentage points annually. The pattern holds in the Defensive Universe, where TSF N20 delivered 7.6% versus 7.7% for TSF N100. TSF timing appears to substitute for diversification by providing an alternative mechanism for risk reduction: rather than spreading exposure across more positions, timing reduces exposure during unfavorable periods. However, extreme concentration (N10, N20) shows inconsistent results, particularly during 2023-2025, suggesting a minimum diversification threshold below which timing cannot fully compensate. Complete signal files are available for independent validation.
TSF timing improves low volatility performance regardless of stock universe. The question addressed here is whether constraining stock selection to defensive sectors (Health Care, Consumer Staples, Utilities, Real Estate) provides additional benefit when combined with TSF timing. The answer reveals a striking interaction effect: defensive sector constraints are a liability without timing but become an advantage with timing. During 2023-2025, defensive sector constraints reduced native returns from 5.8% to 2.8% CAGR—a 3.0 percentage point penalty. TSF timing improved both universes, but the improvement was larger for the defensive universe: Broad TSF delivered 6.9% CAGR while Defensive TSF delivered 8.1% CAGR. The same constraint that cost 3.0 percentage points without timing generated a 1.2 percentage point advantage with timing. This pattern persists across time periods. During 2016-2025, defensive constraints imposed a 1.9 percentage point native penalty but TSF-timed defensive portfolios closed this gap while delivering lower maximum drawdowns. Win rates favor the defensive TSF combination, reaching 81% over the full 20-year sample versus 68% for broad TSF. These results demonstrate that defensive sector selection and TSF timing are complements: timing transforms sector constraints from performance drag to performance advantage. Complete signal files are available to qualified researchers and institutional investors for independent validation.
Low volatility strategies have driven billions in outflows from defensive factor ETFs over the past three years. The core problem: during bull markets, these strategies barely exceed risk-free returns. During 2023-2025, native low volatility strategies using 346 S&P 500 stocks delivered only 5.8-6.1% CAGR—less than 1.1% above T-bills. Investors accepted equity risk for returns barely above a savings account. Temporal Structural Forecasting (TSF) timing signals address this structural limitation. TSF Low Volatility delivered 6.9-7.5% CAGR during 2023-2025, more than doubling the excess return over risk-free rates. Sharpe ratios improved from 0.20-0.23 to 0.31-0.37, and win rates increased from 60% to as high as 77%. Over the full 20-year sample (2006-2025), TSF added 0.8-1.8 percentage points of annual CAGR while reducing maximum drawdowns by 6-8 percentage points. During the 2016-2025 bull market decade, TSF improved CAGR from 6.6-7.4% to 9.3-9.4% with win rates reaching 82%. All results derive from a preregistered study with methodology deposited prior to analysis. Complete signal files are available to qualified researchers and institutional investors for independent validation using their own data and systems.
This study conducts a large-scale empirical test of the long-standing claim that bottom-up (BU) forecasting materially outperforms top-down (TD) forecasting. Using the publicly available WalMart M5 Forecasting – Accuracy Track dataset, daily SKU-level sales and revenue forecasts were generated with three classical univariate models—ARIMA, Holt–Winters Exponential Smoothing (HWES), and Simple Exponential Smoothing (SES)—each applied in both BU and TD forms. Forecast accuracy was evaluated through MAE, sMAPE, and RMSE metrics, aggregated to fixed in-month weekly buckets to mirror real-world retail planning cadences.
The analysis spans 140 independent BU–TD cross-sections comprising 2,520 paired forecasts across models, metrics, and measures. All tests yielded statistically neutral results: mean differences were near zero, hit rates hovered around 50%, and significance frequencies matched random expectation. Both directionality tests (binomial and t-test) and threshold-sensitivity analyses confirmed the absence of any consistent advantage for either aggregation direction. The findings therefore fail to reject the null hypothesis that structural direction has no measurable effect on forecast accuracy.
By providing a comprehensive replication across models, metrics, and measures, this study defines the empirical ceiling of improvement attributable to hierarchical directionality. These results provide a statistically verified baseline for assessing hierarchical and reconciled forecasting methods in future studies and formally close the BU–TD debate under classical forecasting conditions.
Commercial Applications
Demand Forecasting
A fully managed planning service that connects to your existing data and delivers real time decision signals, automated forecasts, and weekly actionable reports. TSF builds a Map of Normal so you can monitor current conditions and know when it's time to hold, grow, or protect.
Equity Timing
Statistical stock price forecasts for Swing and Position traders. TSFStocks forecasts specific limit order prices for each trading day, published one week in advance. Historical validation shows an average of 80% probability of profitability within 120 days.
