XLF PBands Backtest: Profitable, But Risk-Adjusted Evidence Is Weak

Executive Summary

Financials are more cyclical than staples, so the same mean-reversion rule should face a harder regime test. This run covers 2016-01-01 to 2025-01-01 using quantstrat in R, with initial equity of $100,000. I compare it with buy-and-hold, apply base friction, and run a first sequence <span class="lf0-tooltip" tabindex="0" style="border-bottom: 1px dotted currentColor; cursor: help;" title="A deliberate check that makes assumptions harsher to see whether the result survives outside ideal conditions." data-tooltip="A deliberate check that makes assumptions harsher to see whether the result survives outside ideal conditions." aria-label="stress test: A deliberate check that makes assumptions harsher to see whether the result survives outside ideal conditions.">stress test</span>.

Research Card

The Setup

I’m testing a simple mean reversion story in financials: when XLF gets pushed below its lower volatility envelope, does it tend to snap back toward the middle rather than keep trending lower? The logic is that sector ETFs like XLF often overshoot on short-term news, macro surprises, or forced de-risking, and then stabilize once the selling pressure fades.

Results

The backtest did support the mean-reversion hypothesis, but only weakly in economic terms. Final equity was 101,371.63 on a 100,000 start, for net P&L of 1,371.63 and CAGR of 0.15% over 2016-01-01 to 2025-01-01. That is positive, but small.

Statistical Limits And Regime Dependence

The key question is whether the evidence survives sampling uncertainty, clustered trades, and regime shifts. P-values are diagnostics, not trading signals.

Regime dependency: this setup can degrade when volatility expands, gaps dominate, or the sector enters a persistent trend. The forward-degradation rows test stability, not parameter-optimization skill.

Costs And Stress Test

I now run a simple friction pass before drafting the article. For this strategy test, the base case assumes 2.00 bps of <span class="lf0-tooltip" tabindex="0" style="border-bottom: 1px dotted currentColor; cursor: help;" title="The difference between the expected execution price and the actual fill. It usually gets worse in fast or illiquid markets." data-tooltip="The difference between the expected execution price and the actual fill. It usually gets worse in fast or illiquid markets." aria-label="slippage: The difference between the expected execution price and the actual fill. It usually gets worse in fast or illiquid markets.">slippage</span> per side and 1.00 bps of spread cost per side, or 6.00 bps per round trip. That is still a model, but it is a better starting point than pretending fills are free.

Watch the break-even round-trip cost: above that level, the historical edge is gone. Monte Carlo separates edge from path luck; the in/out row is a closed-trade split, not walk-forward optimization.

What Breaks It

The trade P&L distribution is decent but not clean: most trades are small winners, with a cluster around modest positive outcomes, while losses are fewer but noticeably larger in size. That means the edge comes from hit rate more than payoff symmetry, and the average loser still matters. Drawdown, in plain English, is the peak-to-trough loss during a rough stretch before the equity recovers.

Verdict

I set out to test whether a simple lower-band mean-reversion rule on XLF captures short-lived overshoots better than it gets trapped in falling momentum, and the evidence says the idea is plausible but not cleanly robust.

What would change my mind: the same rule needs to survive costs while closing the gap versus buy-and-hold, or prove useful as a timing overlay rather than a standalone strategy.

R Notes

The reproducibility contract is the saved artifact set plus the R scripts that created it. The key JSON/RDS files are recorded in the frontmatter.

# Rebuild the evidence layer for one saved strategy output
source('backtesting_blog_posts/statistical_diagnostics.R')
source('backtesting_blog_posts/walk_forward_diagnostics.R')
source('backtesting_blog_posts/analysis_test_selector.R')

out_dir <- file.path('backtesting_blog_posts', 'outputs', 'strat_pbands_mean_reversion_long_only_xlf')
per_trade <- readRDS(file.path(out_dir, 'per_trade_stats.rds'))
portfolio <- readRDS(file.path(out_dir, 'portfolio_summary.rds'))

diagnostics <- build_statistical_diagnostics(per_trade, portfolio)
walk_forward <- build_walk_forward_summary(per_trade)
test_plan <- build_analysis_test_plan(out_dir)

For the current PBands series, parameter_sensitivity uses a deterministic next-bar rerun helper because native quantstrat parameter clones can fail to propagate osMaxPos limits safely. Future strategy families should call register_parameter_sensitivity_builder() with their own sensitivity builder.

Important: Backtests are simulations. Past performance of a simulated strategy does not guarantee future results. Results do not yet fully account for real-world factors including <span class="lf0-tooltip" tabindex="0" style="border-bottom: 1px dotted currentColor; cursor: help;" title="The gap between the price buyers bid and sellers ask. Crossing that gap is a real trading cost." data-tooltip="The gap between the price buyers bid and sellers ask. Crossing that gap is a real trading cost." aria-label="bid-ask spread: The gap between the price buyers bid and sellers ask. Crossing that gap is a real trading cost.">bid-ask spread</span>, market impact, borrow costs, or taxes. This is for educational purposes only and is not investment advice.