Model-Family Search: The Tuesday No-Go Report

Repository reference: cutebacktests

Abstract

The May 12 search did not find a new promotable model family. That is the headline, and it is useful precisely because it is not flattering.

After the May 8 UOA false start, the research loop widened across high-priority UOA variants, free or cheap source families, put-flow branches, late-session continuation, reversal/fade mechanics, and near-miss combinations. The result was a cleaner map of what did not work. Several rows had positive standalone PnL. None became a basket-first paper candidate.

Question

The Tuesday question was not "can we find a green row?" We already had green rows.

The real question was whether any branch could become additive to the current R45 paper basket while staying options-only, causal, quote-aware, and supportable enough for promotion.

Method

The search used the same promotion philosophy as the rest of the repository:

• standalone PnL is diagnostic, not sufficient;
• the candidate has to improve the active basket;
• residual PnL cannot come from only one or two days;
• PBO and DSR must survive selection pressure;
• option execution evidence must fail closed instead of silently filling from stale or missing quotes.

The main evidence came from the May 12 model-family coverage audit and its branch decision notes.

Evidence

The high-priority UOA branches mostly failed for different reasons:

The secondary family set did not rescue the search. Cboe PCR, FRED rates and credit, SEC/FINRA/CFTC/event-pressure sources, cached breadth, and R8 public/context families all failed active-basket checks.

The near-miss combination check was also decisive. It rescored 66 candidates and scanned 31344 pair/triple combinations. The result was 0 scout-ready and 0 promotion-safe combinations. Some combinations improved daily Sharpe, but the residual PnL was still concentrated on roughly 4 days, often with top-3 residual share near 1.0.

What Worked

What worked was the search discipline. The process did not keep relaxing gates until something passed. It let each branch fail for the specific reason the evidence supported.

That matters because model-family research gets expensive when every near miss stays alive. May 12 narrowed the opportunity set. UOA follow-flow needed a new mechanic, not another timing mask. Public macro and breadth sources looked better as context than as standalone sleeves. Near-miss combinations were not a hidden portfolio solution.

What Failed

The main failure was residual-day support. Many candidates could produce attractive local numbers, but the active-basket contribution came from too few days. That is not a paper-trading edge. It is a concentration problem wearing a strategy label.

The second failure was robustness. Some rows looked positive until March/April or other favorable windows were removed. Others had PBO or DSR evidence that was too weak to justify paper promotion.

The third failure was execution readiness. A branch that cannot preserve quote-strict, options-only fills has to stop before paper, even if the virtual row looks useful.

On Paper Trading

The Tuesday no-go result is also a useful way to explain paper-trading issues that show up in many research stacks. Paper trading should protect the system from weak candidates before they start producing noisy live evidence, and is not just a switch that turns a backtest into simulated orders.

Common issues we see teams encounter:

• a candidate can improve standalone metrics while adding little to the active book;
• residual PnL can come from too few days, making paper results hard to interpret;
• quote-strict rows can become too sparse once live freshness and spread checks are applied;
• different variants can overlap so heavily that paper trading them together only duplicates exposure;
• sparse strategies can spend many sessions doing nothing, which makes progress hard to measure.

The integrated paper trading module we are building is meant to make those cases visible earlier. It will track why a model did not trade, whether a skipped setup was intentional, whether quote checks failed, and whether a candidate is actually additive to the active paper book.

Takeaway

May 12 was a no-go day, but it made the next success possible. The research loop learned that filtering the same sparse UOA row set would not solve the problem, that free-source context was not enough on its own, and that deconcentrating near misses was weaker than finding a genuinely different option expression.

The useful instruction coming out of Tuesday was clear: stop trying to rescue the same families, and search for a new executable option structure with cleaner basket additivity.