GDP-Anchored MFAF

What this paper does

A structured response to three persistent failures in fundamental analysis, validated through a rigorous three-company UK backtest with formal statistical testing.

This paper introduces the GDP-Anchored MFAF, a transparent five-layer system for forecasting company revenue growth. The framework anchors all forecasts to global nominal GDP growth - an exogenous IMF-published series - and applies structured, observable, data-linked adjustments across 35 individual dials. Every dial specifies a named formula, a named Bloomberg data source, distributional scoring thresholds grounded in empirical research, and a cited academic rationale.

We validate through a controlled single-country backtest covering Rightmove PLC, Diploma PLC, and XPS Pensions Group over 2009-2024. The Diebold-Mariano test rejects equal predictive accuracy against the naive GDP-anchor-only benchmark at p = 0.003. Directional accuracy is 89.6% with a 95% CI of [81.2%, 97.9%] - entirely above the 50% random baseline.

The Framework

Five structured layers applied to an exogenous GDP anchor. Click any layer to see the full dial breakdown, weights, data sources, and academic rationale.

Why these five layers, and why in this order

The five layers are not arbitrary. Each captures a distinct, empirically documented source of company-level deviation from GDP growth. The ranking by IC - the order in which they are listed here - reflects which signals are most persistent and least contaminated by short-term noise.

L3 comes first because structural characteristics compound. A company with a wide moat and high ROIC vs WACC will, all else equal, grow faster than GDP for years - and that advantage is observable, documented, and slow to erode. Fama-French (2015) and Mauboussin (2014) both confirm this empirically. Structural signals do not need to be re-estimated every quarter.

L2 comes second because industry dynamics set the ceiling. Even the best company in a structurally declining industry will eventually be dragged down. Klepper (1997) documents the industry S-curve with precision: penetration below 30% predicts fast growth, above 70% predicts slowdown. This is predictable from public data well before it shows up in company earnings.

L1 is ranked third despite being the most macro-visible layer, because the GDP anchor already incorporates aggregate conditions. L1 only adjusts for deviations - where a company's specific geography, FX mix, or monetary environment differs from the global average. In the UK backtest, L1 is held constant across all three companies by design, isolating Layers 2-5 as the source of cross-company variation.

L4 is ranked fourth not because management is unimportant, but because management signals are the most susceptible to narrative contamination. This layer only activates on hard, documented evidence: 8-quarter guidance beat rates, post-acquisition ROIC, insider ownership levels. Qualitative impressions of leadership quality are explicitly excluded.

L5 is last and capped at ±2pp because momentum decays. Jegadeesh and Titman (1993) document that momentum persists 3-12 months, not years. It is the highest-frequency layer, designed to capture near-term estimate revision and beat-streak signals that have not yet been incorporated into structural scores. It should be updated every quarter and discounted heavily beyond 12 months.

Formal Statistical Tests

Four formal tests applied. Two have adequate statistical power at n=48. All results reported with honest interpretation.

Figure 6 - Forecast Bias Analysis. Left: scatter of all 48 observations with OLS fit (R-squared = 0.190, p = 0.002). Right: three identifiable bias sources account for the bulk of systematic error.

What the Scores Mean

The composite score C(i,t) is the weighted average across all five layer scores. Here is how to read it and what different levels imply about the forecast and signal.

The backtest shows that this discipline is useful. 89.6% directional accuracy across 48 company-years, with a confidence interval entirely above the 50% random baseline. The model significantly outperforms a naive GDP-only benchmark at p = 0.003. Those two results hold up formally. Everything else in this paper is evidence in progress.