Stage 3 — Walk-forward FMP validation

A convincing backtest has to reproduce the decision that could have been made on each historical date. StrategyNet rebuilds the signal, risk model, universe, and FMP from the information available then. The positions are frozen before the next return appears. Only after that return is known does it become one observation in the validation record.

This is slower than applying today's model to yesterday's data, but it answers the question that matters: how did the process behave when it did not know what was coming next?

What the animation shows

The registered sleeves enter on the left with their versions and current status. The walk-forward engine advances through rebalance dates in the center. At each step it constructs the FMP, locks the weights, observes the next-period return, and accounts for turnover and costs.

The evidence panel keeps ranking statistics separate from portfolio-return statistics. Rank IC and ICIR describe whether the signal orders securities consistently. Net Sharpe and drawdown describe the return path of the constructed FMP. Coverage, exposure leakage, and correlation with the existing book add different information again.

Finally, explicit gates remove candidates that fail the minimum standard. The survivors are ordered for research review; they have not yet been allocated capital.

Freeze first, measure second

For FMP \(j\) constructed at time \(t\), the next gross return is

\[f_{j,t+1}^{\mathrm{gross}} = h_{j,t}^{\mathsf T}r_{t+1}.\]

One-way turnover is

\[\operatorname{TO}_{j,t} = \frac{1}{2}\lVert h_{j,t}-h_{j,t-1}\rVert_1,\]

so a simple net-return calculation is

\[f_{j,t+1}^{\mathrm{net}} = f_{j,t+1}^{\mathrm{gross}} -\kappa_t\operatorname{TO}_{j,t}.\]

Production cost models can be more detailed, but the principle does not change: a high-turnover signal should not be judged on frictionless returns.

Ranking evidence is not return evidence

The daily rank information coefficient compares the signal or construction weights with subsequent security returns:

\[\operatorname{IC}_{j,t} = \operatorname{corr}_{\mathrm{rank},i} \left(x_{i,j,t},r_{i,t+1}\right).\]

Over a trailing window, the annualized ICIR is

\[\operatorname{ICIR}_{j,t} = \sqrt{252}\, \frac{\overline{\operatorname{IC}}_{j,t}} {s_{IC,j,t}}.\]

ICIR rewards a ranking relationship that is positive and stable. It is not the Sharpe ratio of the FMP's returns. The two can disagree because portfolio construction, covariance, concentration, turnover, and costs sit between a stock ranking and a realized sleeve return.

Promotion gates before ordering

Candidate selection starts with requirements rather than a league table. A deployment may require positive out-of-sample IC and net return, minimum coverage, bounded drawdown and turnover, acceptable exposure leakage, and low enough redundancy with sleeves already selected.

After those gates, a transparent ordering score can combine standardized evidence:

\[q_j = \theta_1 Z(\operatorname{ICIR}_j) +\theta_2 Z(\operatorname{Sharpe}^{\mathrm{net}}_j) -\theta_3 Z(|\operatorname{MDD}_j|) -\theta_4 Z(\operatorname{Turnover}_j) -\theta_5 Z(\operatorname{Redundancy}_j).\]

The coefficients are configurable and should be shown when they are used. The important distinction is that \(q_j\) is a promotion score for ordering research candidates. It is not automatically the expected-return estimate used by the portfolio allocator.

What leaves this stage

The output is an ordered set of FMP sleeves with their evidence attached: return history, IC history, turnover, costs, drawdown, coverage, correlations, and gate results. Stage four can select from that set while still seeing why a sleeve survived and what risks it adds.