How to Backtest a Tax-Loss Harvesting Algorithm with Synthetic Households

What 'realistic' means for a TLH backtest

There are three dimensions of realism that matter for TLH backtesting and that most synthetic fixtures fail to capture.

First, lot-level structure. A real taxable account isn't a single position with an average cost basis; it's a stack of lots, each with its own acquisition date and cost basis, that the TLH algorithm picks among using specific-ID accounting. Test fixtures that aggregate to position-level averages produce TLH yields that are systematically wrong because the harvest opportunity is concentrated in specific underwater lots, not in the position average.

Second, cross-account wash-sale dynamics. The IRS wash-sale rule disallows a loss when substantially identical securities are repurchased within 30 days — including in the spouse's IRA, the same investor's other accounts, or even certain options positions. About 6% of would-be-harvestable lots in a real-world investor's account are blocked by cross-account conflicts. Test fixtures that ignore this dimension over-estimate TLH yield by a similar amount.

Third, QSBS interactions. For founder and early-employee clients, the Section 1202 holding-period and gross-asset-test interact with the TLH decision — selling a QSBS-eligible position before the 5-year mark forfeits the federal capital-gains exclusion. A TLH algorithm that doesn't recognise QSBS positions can recommend harvesting against tax savings of 20-25% while triggering a forfeiture worth 23.8%.

The backtest pipeline

A defensible TLH backtest pipeline has three stages. Each stage produces an artifact you'll show to a stakeholder skeptical that the result is real.

Calibration: why the corpus's deviation rating matters

For backtesting to produce reproducible results, the test corpus must be deterministic. Two backtest runs of the same algorithm against the same corpus snapshot must produce the same harvest schedule, the same constraint filter, and the same yield calculation. This is the rationale for the 'Minimal' deviation rating on the WealthSchema TLH bundle — lots and balances are deterministic given the seed, so regression tests work.

This property is what lets you make claims like 'the new algorithm produces 12% higher TLH yield than the prior version' with statistical confidence. Without determinism, the comparison drowns in noise.

Stress regimes: backtesting beyond the snapshot

The corpus is calibrated to a 'normal' market environment — about 18% of lots show net unrealised losses at snapshot. For backtesting under stress regimes (a 2022-style downturn, a 2008-style crisis), apply a parameterised re-pricing transformation to the corpus before running the backtest. The Methodology PDF includes the methodology for stress-regime re-pricing; the corpus structure preserves the lot-level integrity through the transformation.

// Stress-regime re-pricing pseudocode
function applyStressRegime(corpus, regimeName) {
  const factors = STRESS_FACTORS[regimeName];
  return corpus.map(household => ({
    ...household,
    accounts: {
      ...household.accounts,
      taxable: household.accounts.taxable.map(account => ({
        ...account,
        lots: account.lots.map(lot => ({
          ...lot,
          current_price: lot.current_price *
            factors[lot.asset_class] *
            (1 + lot.beta * factors.market_excess)
        }))
      }))
    }
  }));
}

Reporting: what the backtest result actually says

The backtest output isn't 'the algorithm produces X bps of TLH yield.' It's 'the algorithm produces X bps of TLH yield against Y corpus under Z assumptions.' The corpus citation matters because the answer changes meaningfully across corpora — a corpus weighted toward ETF-heavy holdings produces lower TLH yield than a corpus weighted toward single-stock holdings, even with the same algorithm. The assumption citation matters because tax-rate assumptions, state-tax assumptions, and short-vs-long-term gain composition all affect the dollar yield even with the same harvest schedule.

Report the backtest with the full provenance: corpus version (e.g. 'WealthSchema TLH bundle, 2026 annual refresh, 350 households'), constraint set (e.g. 'wash-sale within 30 days across all accounts in household graph'), and assumption set (e.g. 'federal marginal rate 37%, state marginal rate 9.3% for CA residents, 0% for FL residents'). The provenance is what makes the backtest defensible to investment committees, regulators, and customers.

Key takeaways

• TLH backtests fail when the test fixture lacks lot-level structure, cross-account wash-sale dynamics, or QSBS interactions — three dimensions most synthetic fixtures skip.
• A defensible backtest pipeline has three stages: snapshot run, constraint filter, yield calculation. Each stage produces a reviewable artifact.
• Determinism (the 'Minimal' deviation rating) is what makes regression testing work — same corpus + same algorithm always produces same yield.
• Stress-regime backtesting works by parameterised re-pricing applied to the corpus snapshot, preserving lot-level integrity through the transformation.
• Backtest results should always cite corpus version, constraint set, and assumption set — the provenance is what makes the result defensible.

FAQ