Skip to content
TensorStreet Labs
Built-in intelligence

Model Zoo

Start with a model. Build on the evidence.

Walk-forward research

Meet TensorStreet’s research model

Inside APS-MRCR

Multi-Resolution Convolutional-Recurrent network

Two stocks can finish a month with the same return after very different daily journeys. APS-MRCR learns from those journeys, their market context and the monthly picture.

A neural model built to rank stocks.

It is trained to order stocks by their next-month return. You turn that ranking into a portfolio, then test whether the signal adds useful information beyond a simple rule like momentum.

The output is a relative ranking score, not a price target or a probability of profit.

Three views of a stock. One signal.

Explore what each branch contributes.

60 trading days · 18 daily features

How did the stock get here?

A steady climb, a sharp reversal and a volatile recovery can end with the same return. This branch learns from the sequence of daily price moves, trading volume and technical features.

A CNN detects short patterns over 3 and 5 days. A GRU reads the sequence to learn how those patterns unfold over time.

Combine the views, accounting for uncertainty

Each branch produces a score and an uncertainty estimate. A learned gate combines them for each stock, reducing a branch’s influence when it reports greater uncertainty.

Three independently trained networks are averaged → one stock-ranking signal.

From the research manuscript · Table 1

Does combining the views help?

The combined model showed stronger agreement with subsequent stock rankings than either of these simpler versions in the paper’s test period.

Paper-reported rank correlation with next-month returns, January 2017 to June 2026
Model inputsRank IC
Daily prices only0.010
Monthly inputs only0.024
APS-MRCR · combined0.032

Rank IC measures agreement between the model’s ordering and realised returns. It is neither a return percentage nor a hit rate.

Learn from the past. Evaluate on later months.

In the research, the model is refit each January using earlier data. The most recent 12 training months select the stopping point; subsequent months are used for evaluation.

Historical study: Jan 2017–Jun 2026, using current S&P 500 members (survivorship bias); June is partial. These are manuscript results, not live performance or a forecast for your strategy.

For researchers: architecture, training and what remains to test

APS means AI Pattern Score. MRCR combines two CNN–GRU sequence encoders with a monthly MLP and an uncertainty-aware gate. The paper specifies 93,114 parameters and a three-seed ensemble. Daily inputs are observed through the previous month’s close; the target is the within-month rank of next-month returns.

The paper reports tests against momentum, linear and tree-based signals. Point-in-time index membership, an independent future-period replication, and a separate test of the uncertainty gate’s contribution remain open. The platform reuses committed historical predictions; opening a strategy does not train or freshly score the model.

Source: APS: Multi-Resolution Neural and Linear Price-Path Signals Beyond Momentum in Large U.S. Stocks, research manuscript, architecture and training section and Table 1. The reported comparisons describe that historical study; your portfolio settings can produce different results.

Research models — models