A stock-selection table with only momentum looks sparse. Would adding quality, low volatility, and trend strength make it safer? Before averaging columns, check whether two merely repeat earlier price strength and whether another rewards the opposite direction from the one intended.
Even correctly calculated signals can combine without changing a single holding. The task is to identify what an added column changes and whether that difference warrants more research, rather than fill the table with indicator names.
Define the proposed addition
You will align direction and scale, hand-calculate a simple combination, compare single and combined signals under consistent conditions, and preserve a valid conclusion of no demonstrated increment. Complex models receive entry conditions here; you will not train one.
Price and financial characteristics can enter factor strategies, but names cannot replace a mechanism and implementation. CFA Institute's overview connects a thesis, data processing, testing, evaluation, and portfolio risk and cost.[1] Adding a signal starts with a testable explanation.
Momentum asks whether earlier relative strength continues. A second price feature might prefer smaller recent daily movements, asking whether it distinguishes similarly strong stocks. This is a hypothesis: quiet stocks can still gap, and strength and quietness can reflect the same industry environment.
Keep G84's distinction between signal treatment and portfolio-exposure constraints. Retain raw momentum in this exercise; do not silently substitute industry residuals. Two combined columns do not automatically create industry-neutral or diversified holdings.
Align direction before combining units
The standalone C7-85 exercise reuses A, B, C, and E. D remains ineligible because of the missing momentum endpoint. After m closes, momentum is the then-available total-return index P(m−1)/P(m−12) minus one, skipping m. Other eligibility, stable IDs, and next-execution timing retain their original definitions.
The new inputs are synthetic close-to-close total returns for the last two normal trading days, both known before the decision. Define Q as their average absolute return. Lower Q receives greater preference. Two days keep arithmetic visible; this is not a reliable general risk-estimation window or a recommended live parameter.
| ID | Momentum M | Earlier daily return | Latest daily return | Q: mean absolute daily return |
|---|---|---|---|---|
| A | 10% | −1% | 1% | 1% |
| B | 20% | −2% | 2% | 2% |
| C | 20% | −4% | 4% | 4% |
| E | −10% | −3% | 3% | 3% |
Q is not sample standard deviation. It measures absolute movements from zero. A value of 2% is not a forecast of 2% future volatility. It uses no next-holding-period return. Real research would separately register sufficient history, window length, valid observations, and abnormal-price handling, then obtain corresponding data.
A direct average of M and Q rewards higher Q, contrary to the intended preference. Negating Q fixes direction but leaves scale choices: inconsistent use of 20 versus 0.20 can alter relative influence. Equal weight needs a definition of which transformed scores receive that weight.
Mean/standard-deviation scaling is common and can be sensitive to outliers.[2] It is not the only choice. Predefine cross-sectional rank scores here, with higher always meaning more preferred. No standard-deviation scaling is used in this version.
Preserve equal signals before resolving a trade tie
For four stocks, the least preferred rank is one and the most preferred four. Equal values receive the average of occupied ranks. Momentum's ascending order is E, A, then B/C, giving B and C 3.5 each. Q preference reverses its raw direction: C, E, B, A.
Each score is its preference rank minus one, divided by three. Scores lie between zero and one. The combination averages the two scores. Ranking discards original gap sizes: momentum's ten-percentage-point difference no longer enters as a ten-point gap. This is an explicit tradeoff, not a free improvement.
| ID | M rank | Q preference rank | M score | Q score | Equal-weight combination |
|---|---|---|---|---|---|
| A | 2 | 4 | 1/3 | 1 | 2/3 |
| B | 3.5 | 3 | 5/6 | 2/3 | 3/4 |
| C | 3.5 | 1 | 5/6 | 0 | 5/12 |
| E | 1 | 2 | 0 | 1/3 | 1/6 |
B's average is (5/6 plus 2/3)/2, or 3/4. The 50/50 weights apply to standardized signals, not separate halves of an account. First calculate one combined score for each stock, then construct one portfolio. Combining complete strategies requires their separate rules, holdings, and accounts, developed later in G91.
Continue selecting the top 20%, rounded upward: one stock. Only a tie in the final combined score invokes stable-ID trading priority. Momentum alone selects B; Q alone selects A; the equal-weight combination still selects B. Changed scores need not change holdings.
A second name can give old information another vote
Create a column called Enhanced Trend, defined as twice M. Numbers differ but ranks are identical. Averaging the two rank scores reproduces the original momentum score on every row. The new label adds no information.
Starting with 50/50 momentum and Q, adding this duplicate with three equal column weights changes the family allocation. Momentum-related information receives two thirds and Q one third. What looks like an additional source has increased one viewpoint's implicit weight.
Real redundancy is not always an exact copy. During development, inspect signal correlations over time, missing coverage, and overlapping holdings. Different windows or formulas can express much of the same price move. High correlation need not force deletion; low linear correlation does not establish complementarity, since a nonlinear transformation may preserve similar information.
Return to economic exposures as well. Statistically different columns may both select the same industry, size, or liquidity conditions. Correlation checks locate repetition; they neither replace G84's exposure work nor prove diversified returns.
Compare the addition under shared conditions
An ablation removes one component at a time to examine the change. Preserve at least four versions: momentum alone, Q alone, equal combination, and momentum combined with its duplicate. Do not report only the attractive row.
Use the already-revealed C7-82 gross total returns, including distributions, from the first normal open in m+1 through that in m+2: A 4%, B −2%, C 6%, E 0%. Full fills are assumed. The same-pool equal-weight benchmark earns 2%. These outcomes appeared earlier, so this is retrospective mechanism checking, not an unseen holdout test.
| Version | Selection | Gross holding-period return | Active difference versus pool |
|---|---|---|---|
| Momentum alone | B | −2% | −4 percentage points |
| Q alone | A | 4% | 2 percentage points |
| Equal M/Q | B | −2% | −4 percentage points |
| Equal M/duplicate M | B | −2% | −4 percentage points |
The equal combination adds neither a holding nor a gross-return increment over momentum in this period. The duplicate supplies no new information. Q alone performs better here without establishing persistent prediction. The best displayed row cannot be retroactively declared the original plan.
Real comparisons need consistent data versions, a common eligible universe, execution and return intervals, benchmark, cost scenarios, and research-attempt budget. If the new field is missing for a stock, rerun momentum within the same complete-case pool and report excluded coverage separately. Otherwise apparent improvement could come from changing the universe.
No complete multiperiod fills and cash ledger are supplied, so this table reports gross returns. A ranking improvement may induce more frequent changes and higher costs. Identical fee rates do not produce identical dollar costs across versions. Net difference, turnover, drawdown, and exposures need separate comparisons before implementation can be accepted.
Raising Q's weight is another experiment
After seeing A's 4%, increasing Q to 75% becomes tempting. Calculate it if useful, while recording that an observed outcome prompted the choice. It is development, not independent confirmation of the frozen equal-weight version. Repeated trials followed by winner-only reporting are central concerns in research on backtest overfitting.[3]
Weights, directions, windows, missing values, ranks, and outlier treatments are all choices. Define development, validation, and final evaluation roles beforehand, limit candidates, and preserve every attempt. Once a final result drives a modification, record the inspection and contamination; a new name cannot make the evidence unseen again.
Ranking an allowed contemporary cross-section itself needs no future return. Learning historical means, imputation values, or model parameters does require restricting estimation to permitted training inputs. Apply frozen transformations afterward; do not estimate them from the complete future sample and move them backward.[4]
A complex model needs a specific entry reason: what relation does the simple baseline miss, how will the addition be evaluated fairly, are data and cost adequate, and who can explain and maintain it? An algorithm name supplies more choices, not incremental evidence.
Rebuild the judgment under a changed condition
Explanation question: Why does adding 2M not add information?
Application question: B's Q is missing. An analyst assigns it zero combination score but retains B in the momentum baseline, then reports improvement. Identify the problems before checking.
M and 2M have identical rank scores, so their average is unchanged. Duplication can also increase their family's hidden weight. Missing Q is not low preference; zero changes its meaning. Under a predefined complete-case policy, exclude B from both versions and recalculate ranks, selection count, and benchmark, reporting coverage. Other imputation policies need prior definition and evaluation.
Actually rebuild that common pool. A, C, E remain, so score denominator becomes two. Their momentum scores are 1/2, 1, 0; Q preference scores are 1, 0, 1/2. Equal combinations become 3/4, 1/2, 1/4, selecting A. Momentum alone selects C.
The common-pool benchmark is (4% plus 6% plus 0%)/3, about 3.3333%. Combination earns 4%, below momentum's 6%. Finding an apparently decent combined selection after excluding a missing field does not demonstrate an increment. Both comparison and benchmark must change with the pool.
Change the signal weights
In a separate variation, retain all original four-stock scores but weight momentum 25% and Q 75%. Before consulting returns, calculate A and B, then identify the first choice. Can its higher table return establish a validated optimal weight?
A gives 1/12 plus 3/4, or 5/6. B gives 5/24 plus 1/2, or 17/24. C is 5/24 and E 1/4, so A is selected. Its supplied period returns 4%, but the exercise uses already-seen outcomes and must be logged as retrospective.
An improvement claim requires a separately frozen plan and suitable evidence that did not guide selection. Preserve the losing equal-weight row instead of deleting it. If you already knew the future labels, record that fact even when recalculating scores before looking back at the return table.
Save a combination experiment card
Retain both signals' fields, mechanism, direction, observation window, availability, missing policy, score formula, weights, and choice rationale. State which evidence was already seen. Preserve the proposal before revealing evaluation outcomes; a proposal reconstructed now is retrospective.
Attach raw rows, processed scores, combination, eligibility, rankings, holdings, full ablations, coverage changes, benchmark, cost status, and all attempts. Hand-check two randomly chosen rows. Look for duplicate-family weighting, mismatched pools, and missing no-increment versions.
Ask an AI or CoTrading for this calculation chain if using it. If the current tool cannot export the needed fields or has not actually run, retain hand calculations and the unverified list. Real C9 data, supplier selection, and experiments remain unfinished. Four synthetic rows cannot certify real multifactor research.
Ask what the new column changes
Check mechanism, direction, scale, and missing values before redundancy. The practical question is how the column changes eligibility, ranking, holdings, and common exposure. A weight label alone does not answer it.
Keep single and combined versions under shared conditions, including failures. No demonstrated increment can be a useful research finding. Outcome-driven weight changes open another honestly logged experiment; they do not turn the past into independent evidence.
References
- CFA Institute: Active Equity Investing—Strategies. Public 2026 process overview; paid curriculum not read.
- scikit-learn: Preprocessing Data. Scaling and outlier sensitivity; the rank score here is the book's own definition.
- Bailey et al.: The Probability of Backtest Overfitting. February 2015 revision; introduction on repeated trials and selection, not a PBO calculation for this exercise.
- scikit-learn: Common Pitfalls. Training/test separation and preprocessing leakage.
Sources checked October 30, 2026. Added daily returns, scores, and weight variations are synthetic mechanism exercises, not real factor-effectiveness evidence. AI-assisted draft; independent AI full-text and bilingual review is complete. Human financial review and target-reader testing have not been conducted and are not scheduled for this production round.