What the backtests taught me

Lower cost is arithmetic. A better investment is a separate question.

A fill below the position’s immediate pre-fill average reduces its average cost. A fill above it raises that average, including a manual purchase. Neither tells you what the investment is worth. A cheaper entry can still lose money.

The account result has to include unused cash as well as positions. A comparison with an old average-cost reference is an explanation of an entry, not extra profit to add to the account’s return.

What was tested

The study tested 40 fixed-dollar and remaining-cash percentage rules against investing upfront, calendar purchases (DCA), and cash. Fifteen symbols had enough history for 252-session windows, using adjusted daily Alpaca bars from January 2016 through September 4, 2026.

Each run started with $1,000. Automated rules opened with a $1 seed; DCA made 12 equal purchases every 21 sessions. Daily closing signals traded at the next adjusted open. The base model applied 10 basis points of adverse pricing on purchases and final liquidation, with zero cash yield. It excluded taxes, subscription fees, intermediate sells, and discretionary manual trades.

There were 282 development windows ending before 2023 and 164 later windows starting in 2023 or later. Windows crossing the boundary were excluded. The 19,178 runs are not independent observations: windows overlap and symbols share market exposure.

The strongest development setting did not travel cleanly

Buying with 20% of remaining deployable cash at 1% below average cost had the highest median paired excess return versus investing upfront in development: +2.65 percentage points, compared with +0.13 points in the later sample. This setting was selected retrospectively. The later sample is not prospective validation or a recommended default.

Only 1 of 40 automated settings had a positive median paired difference versus investing upfront in the later sample; 4 of 40 did versus DCA. Changing the asset universe or window selection also changed the result:

Broad ETFs are SPY, QQQ, IWM, XLF, and XLU. Non-overlapping windows were selected greedily within each symbol and split, with the next start strictly after the previous selected end. Cross-symbol dependence remains. Paired differences compare each rule with its benchmark in the same window before taking the median; they are not differences between separate median returns.

Across all 164 later windows, this rule’s median cash-inclusive return was 13.64%, with median ending cash of $167.62. Upfront investing returned a median 27.87%; DCA returned 13.50%. These separately computed medians answer a different question from the paired comparisons above.

Waiting has a cost. Buying has a risk.

In the synthetic rally, the 20% / 1% rule returned 0.06% while upfront investing returned 59.68%. In the permanent-decline illustration, that same rule lost 78.94%. These are constructed paths, not observed market returns or forecasts. They show why neither waiting nor buying more is a universal answer.

A price can remain below the updated trigger after a purchase. Another eligible check can buy again without a fresh decline. Percentage sizing makes later orders smaller, but does not preserve a fixed cash reserve. Position-value caps can also reopen room as prices fall; they are not cumulative spending limits.

Tests with different costs, hypothetical cash yield, and matched reserves changed the comparisons. Costs also changed the average-cost reference and later triggers, so relative results did not always worsen as modeled costs rose. That is not a reason to pay higher costs.

What this study cannot tell us

This was a daily, single-security study of the personal app used to refine Position Builder, called MCT in the report. It was not a replay of the website’s exact engine, actual microfolio results, or live brokerage fills. It did not model shared cash competition, scan priority, partial fills, actual limit execution, all partner fees, or taxes. The current website also supports optional exits that were absent from the historical grid.

The selected holdings introduce selection and survivorship bias. Adjusted bars approximate total returns using synthetic units, rather than literal broker shares or tax lots. The retrospective split and dependent samples limit what can be inferred. Paper trading is useful for testing operations; it does not prove net live performance.

The next performance test would need a policy fixed in advance, a point-in-time universe, actual order and cashflow records, multi-position cash competition, and all costs. Whether the product improves discipline, keeps users within their intended commitments, or is worth paying for remains untested.