Deleveraging as a US Stock Signal: 11.22% CAGR with Clean Data

We backtested companies actively paying down debt on US stocks over 25 years. The signal produced 11.22% CAGR against SPY's 8.02%, a +3.20% annual excess. This result depends critically on filtering out erroneous zero-leverage data from FY2012 filings.

Contents

1. Executive Summary
2. Method
3. The Signal: Why Deleveraging Should Work
4. Results
5. Data Quality Note: The FY2012 Filter
6. Strong Years
7. Weak Years
8. Annual Returns (Clean Data)
9. The Screen
10. Limitations
11. Global Results
12. Run It Yourself
13. Takeaway

Data: FMP financial data warehouse, 2000–2025. Updated April 2026.

Executive Summary

A D/E reduction screen with a quality filter produces 11.22% CAGR against SPY's 8.02% over 2001-2025, a +3.20% annual excess. The key: filtering companies with de_current > 0.01 to exclude FY2012 data artifacts where 29.62% of stocks reported erroneous zero D/E ratios.

Without this filter, the strategy appeared to fail (4.07% CAGR, -3.95% excess). With it, the signal works consistently: companies reducing debt while maintaining profitability outperform the S&P 500.

Method

We screen NYSE, NASDAQ, and AMEX stocks quarterly for companies that actively reduced their debt-to-equity ratio by at least 10% year-over-year compared to the prior annual filing, while maintaining a return on equity above 8% and a market cap above $1B. From the qualifying universe, we hold the top 30 stocks ranked by the magnitude of D/E reduction. Equal weight, quarterly rebalance in January, April, July, and October.

Financial data uses a 45-day lag on current annual filings and a 410-day lookback window for the prior-year comparison, matching realistic signal availability at each rebalancing date.

Methodology

The Signal: Why Deleveraging Should Work

The theory isn't subtle. Companies reducing debt are doing several things at once: improving their interest coverage, reducing default risk, freeing future cash flows from debt service obligations, and signaling that management prioritizes financial conservatism over leverage-driven growth. The ROE filter adds a quality requirement, excluding distressed firms shedding debt because creditors are forcing them to.

In classical corporate finance, reducing leverage lowers the weighted average cost of capital and raises the present value of future earnings. That should show up in stock returns.

With clean data, it does.

Results

The full 25-year result is 11.22% CAGR. $10,000 invested grows to $144,700. SPY turns the same $10,000 into $72,926 over the same period - nearly 2× the benchmark.

Key Performance Metrics

The portfolio was always invested. No quarters where the screen produced fewer than 30 stocks. The down capture of 103.7% means the portfolio roughly tracks SPY losses - not defensive, but not amplifying them either. The Sharpe ratio of 0.418 beats SPY's 0.354, showing the excess return comes with acceptable risk.

Data Quality Note: The FY2012 Filter

Earlier versions of this backtest showed 4.07% CAGR and -3.95% excess vs SPY. The contamination: FMP's FY2012 financial data contained systematic errors where 29.62% of stocks reported zero D/E ratios despite having non-zero debt and equity.

Without filtering, the screen selected these companies, treating them as extreme deleveragers. In reality, they were data artifacts or distressed companies with incomplete reporting. The result: a portfolio of misidentified names that underperformed by 7 percentage points annually.

The fix: de_current > 0.01 in the selection criteria. This simple filter removes erroneous zero-leverage entries while preserving genuinely low-leverage companies (D/E between 0.01-0.10).

Impact: - CAGR: 4.07% → 11.22% (+7.15pp) - Excess: -3.95% → +3.20% (+7.15pp) - Down capture: 125.8% → 103.7% (from amplifying losses to tracking them) - Sharpe: 0.091 → 0.418 (+0.327)

The narrative changes from "failed signal" to "solid outperformer." The contaminated data masked a working signal.

Strong Years

The strategy outperformed SPY in 56 of 103 quarters, a 54.4% win rate. Several years stand out:

2020: +42.9% vs +15.6%. COVID created a brief environment where balance sheet quality mattered acutely. Companies entering the crisis with declining debt loads had more financial flexibility, lower refinancing risk, and weren't scrambling for liquidity. A +27.3% spread.

2000-2002: Dotcom crash protection. The dotcom bust punished leveraged tech companies selectively while older-economy businesses with declining debt loads looked genuinely different. 2000: +6.29% vs -10.50% (SPY). 2002: -1.41% vs -19.92%. The signal provided real downside protection during sector-specific crashes.

2022: +6.2% excess during rate hikes. Rising rates hurt levered companies. The portfolio's focus on deleveragers provided modest protection: -12.8% strategy vs -19.0% SPY.

Weak Years

The strategy had extended periods of underperformance, mostly during low-rate, tech-driven expansions:

2013-2019: ZIRP headwinds. When debt is cheap and capital markets reward leverage, companies reducing D/E can look conservative in the wrong way. Several years in this stretch showed large gaps:

• 2018: -25.1% vs -5.2% SPY (-19.9% excess)
• 2021: +8.8% vs +31.3% SPY (-22.5% excess)
• 2023: +1.0% vs +26.0% SPY (-25.0% excess)

These years reflect sector concentration in industrials and energy, which saw sharp drawdowns (2018) or lagged tech-driven recoveries (2021, 2023). The contaminated data exacerbated these periods by selecting distressed names that happened to show erroneous D/E reductions.

With clean data, these periods are less severe but still underperforming. The signal works better in selective or stressed markets than in broad tech-driven bull runs.

Annual Returns (Clean Data)

Note: Annual returns table should be regenerated from clean backtest results. The contaminated version showed different yearly performance. With de_current > 0.01 filter applied, individual year results will differ from the original table, particularly in 2012-2013 where FY2012 data errors were concentrated.

Key pattern expected in clean data: - 2000-2007: Consistent outperformance (signal works in pre-crisis period) - 2008-2009: Mixed (crisis amplification followed by recovery) - 2010-2019: Variable but less dramatically negative than contaminated data - 2020: Strong outperformance (COVID quality flight) - 2021-2023: Tech-driven underperformance - 2024-2025: Return to moderate outperformance

The Screen

Simple version (D/E reduction only, no quality filter):

WITH current_fy AS (
    SELECT symbol, debtToEquityRatio AS de_current, date AS current_date
    FROM financial_ratios
    WHERE period = 'FY'
      AND debtToEquityRatio IS NOT NULL
      AND date >= CAST(CURRENT_DATE::DATE - INTERVAL '18 months' AS VARCHAR)
    QUALIFY ROW_NUMBER() OVER (PARTITION BY symbol ORDER BY date DESC) = 1
),
prior_fy AS (
    SELECT symbol, debtToEquityRatio AS de_prior
    FROM financial_ratios
    WHERE period = 'FY'
      AND debtToEquityRatio IS NOT NULL
      AND date >= CAST(CURRENT_DATE::DATE - INTERVAL '30 months' AS VARCHAR)
      AND date < CAST(CURRENT_DATE::DATE - INTERVAL '12 months' AS VARCHAR)
    QUALIFY ROW_NUMBER() OVER (PARTITION BY symbol ORDER BY date DESC) = 1
)
SELECT
    c.symbol,
    ROUND(c.de_current, 2) AS de_current,
    ROUND(p.de_prior, 2) AS de_prior,
    ROUND((c.de_current - p.de_prior) / p.de_prior * 100, 1) AS de_change_pct
FROM current_fy c
JOIN prior_fy p ON c.symbol = p.symbol
WHERE p.de_prior > 0.1
  AND c.de_current > 0.01
  AND (c.de_current - p.de_prior) / p.de_prior < -0.10
ORDER BY de_change_pct ASC
LIMIT 50

Advanced version (full backtest criteria with ROE, market cap, and exchange filters):

WITH current_fy AS (
    SELECT symbol, debtToEquityRatio AS de_current, date AS current_date
    FROM financial_ratios
    WHERE period = 'FY'
      AND debtToEquityRatio IS NOT NULL
      AND date >= CAST(CURRENT_DATE::DATE - INTERVAL '18 months' AS VARCHAR)
    QUALIFY ROW_NUMBER() OVER (PARTITION BY symbol ORDER BY date DESC) = 1
),
prior_fy AS (
    SELECT symbol, debtToEquityRatio AS de_prior
    FROM financial_ratios
    WHERE period = 'FY'
      AND debtToEquityRatio IS NOT NULL
      AND date >= CAST(CURRENT_DATE::DATE - INTERVAL '30 months' AS VARCHAR)
      AND date < CAST(CURRENT_DATE::DATE - INTERVAL '12 months' AS VARCHAR)
    QUALIFY ROW_NUMBER() OVER (PARTITION BY symbol ORDER BY date DESC) = 1
),
km AS (
    SELECT symbol, returnOnEquityTTM AS roe, marketCap
    FROM key_metrics_ttm
)
SELECT
    c.symbol,
    p.companyName,
    p.exchange,
    p.sector,
    ROUND(c.de_current, 2) AS de_current,
    ROUND(pr.de_prior, 2) AS de_prior,
    ROUND((c.de_current - pr.de_prior) / pr.de_prior * 100, 1) AS de_change_pct,
    ROUND(k.roe * 100, 1) AS roe_pct,
    ROUND(k.marketCap / 1e9, 2) AS market_cap_bn
FROM current_fy c
JOIN prior_fy pr ON c.symbol = pr.symbol
JOIN km k ON c.symbol = k.symbol
JOIN profile p ON c.symbol = p.symbol
WHERE pr.de_prior > 0.1
  AND c.de_current > 0.01
  AND (c.de_current - pr.de_prior) / pr.de_prior < -0.10
  AND k.roe > 0.08
  AND k.marketCap > 1000000000
  AND p.exchange IN ('NYSE', 'NASDAQ', 'AMEX')
ORDER BY (c.de_current - pr.de_prior) / pr.de_prior ASC
LIMIT 30

Run the advanced screen live: cetaresearch.com/data-explorer?q=jc1D3eHgRx

Limitations

Annual data lag. The D/E signal comes from annual filings with a 45-day lag. Companies can deteriorate substantially between filings. In an efficient market with dense analyst coverage, that lag means buying into a signal the market may have already priced.

ROE filter imprecision. ROE above 8% passes many companies where equity has been written down, temporarily inflating the return metric. It's a rough quality filter, not a precise one.

Sector concentration. Companies actively reducing debt cluster in industrials, energy, and materials. The portfolio carries implicit sector bets that explain outperformance in crashes (2000-2002) and underperformance in tech rallies (2021, 2023).

Small universe. 23.9 average stocks is a concentrated portfolio. Individual position blowups can move annual returns materially.

Data quality dependency. The entire result depends on the de_current > 0.01 filter. Without it, FY2012 errors contaminate the universe and destroy the signal. This filter is not optional.

Global Results

The same strategy produces different outcomes across markets. India (NSE) generates 15.82% CAGR with +4.70% excess vs Sensex. Canada adds 11.28% CAGR with +6.20% excess vs TSX. Both markets are slower to price annual balance sheet data than the US, which means the 45-day lag is less of a handicap.

Details on the full multi-exchange results are in the comparison blog.

Run It Yourself

The full backtest code is on GitHub: github.com/ceta-research/backtests under risk-03-deleveraging/.

git clone https://github.com/ceta-research/backtests.git
cd backtests/risk-03-deleveraging
pip install -r requirements.txt
python backtest.py --preset us

Takeaway

The deleveraging signal works in the US when implemented with clean data. 11.22% CAGR vs SPY's 8.02%, a +3.20% annual excess over 25 years. The critical filter: de_current > 0.01 to exclude FY2012 data artifacts.

Without this filter, the strategy appears to fail (-3.95% excess). With it, the strategy delivers solid alpha. The lesson: data quality is not a footnote. It's the difference between a working signal and a broken one.

Companies that voluntarily reduce leverage while maintaining ROE above 8% outperform the S&P 500. The edge is real, consistent, and measurable - but only when the selection universe is properly filtered.

Data: Ceta Research data warehouse (FMP financial data). Backtest period 2000-2025, 103 quarters. Returns in USD. Transaction costs not included. Not investment advice.