← All tasks
T11 · Factor Investing

Mimicking Finance: Predicting Mutual Fund Manager Behavior Using Deep Learning

Lauren Cohen et al. (3 authors), 2026 · NBER Working Paper 34849

0%
pass@1 across 16 systems
0/16
systems with ≥1 successful attempt
0/5
no system produced a successful attempt
80 attempts · 16 systems × 5
Claude Code + Opus 5
0%
Claude Code + Opus 4.8
0%
Claude Code + Opus 4.7
0%
Codex + GPT-5.6-sol
0%
Codex + GPT-5.5
0%
OpenCode + Kimi-K3
0%
OpenCode + Opus 4.7
0%
OpenCode + GPT-5.5
0%
OpenCode + Sonnet 4.6
0%
OpenCode + Qwen3.7-Max
0%
OpenCode + GLM-5.2
0%
OpenCode + DeepSeek V4-Pro
0%
OpenCode + GLM-5.1
0%
OpenCode + Qwen 3.6 Plus
0%
OpenCode + MiniMax-M3
0%
OpenCode + MiniMax M2.7
0%
all checkers pass Type-A clean, Type-B fail fail checking timeout no result / incomplete
Strategy

What the paper asks for.

Per-manager LSTMs trained on historical 13F holdings can identify "predictable" stocks; cross-sectionally rank stocks by average prediction accuracy and go long the most-predictable quintile / short the least predictable, equal-weighted, rebalanced quarterly.

Type
hybrid
Implementation
pandas
Rebalance
quarterly
Long-short
yes
Declared task setup
Data
One-command fetcher for WRDS-sourced task data
Operational scope
The paper-defined methodological core is retained; the fixed instruction instantiates only the data mappings, interfaces, and evaluation conventions needed for execution.
Six-stage quantitative strategy pipeline
  1. 01 Data loading
  2. 02 Feature engineering
  3. 03 Model fitting
  4. 04 Signal generation
  5. 05 Portfolio construction
  6. 06 Execution
Checker suite

20 properties · 80 attempts.

Each row is one requirement the strategy must satisfy. The bar reports the share of observed checker verdicts that pass; missing checker output is not converted into a property verdict.

TYPE A · DOMAIN-LEVEL (10)
  • A1
    no backfill
    73/74
  • A2
    no global normalize
    74/74
  • A3
    no future data access
    74/74
  • A4
    causal consistency
    74/74
  • A5
    no survivorship bias
    68/74
  • A6
    train test temporal order
    74/74
  • A7
    data frequency alignment
    68/74
  • A8
    transaction timing
    43/74
  • A9
    signal delay test
    74/74
  • A10
    e2e metric consistency
    2/74
TYPE B · PAPER-SPECIFIC (10)
  • B1
    n quintile portfolios
    §Table XIII / §5 “stocks are ranked by the average prediction accuracy of all funds holding that stock and sorted into five equal-weighted portfolios.”
    provenance · paper-explicit
    stage · signal gen
    65/74
  • B2
    long q1 short q5 direction
    §§5 “stocks in which the behavior of fund managers are least predictable (Q1) strongly outperform stocks in which the behavior of fund managers are most predictable (Q5) in the coming quarter.”
    provenance · paper-explicit
    stage · signal gen
    50/74
  • B3
    label classification dead band
    §§3.2 “Δsh_{i,t} = sh_{i,t+1} - sh_{i,t} / sh_{i,t} + 1 ... Y_{i,t} = -1 if Δsh ≤ -0.01, +1 if Δsh ≥ +0.01, 0 if |Δsh| < 0.01. ... A ±1% band around zero counts as 'no change'.”
    provenance · paper-explicit
    stage · feature eng
    65/74
  • B4
    single layer lstm architecture
    §§3.3 “We estimate a single-layer LSTM network on these sequences.”
    provenance · paper-explicit
    stage · model fit
    4/74
  • B5
    manager filter seven years ten securities
    §§3.1 “Funds must span at least seven calendar years and hold at least 10 securities per quarter; otherwise they are excluded.”
    provenance · paper-explicit
    stage · data loading
    67/74
  • B6
    lstm sequence length eight quarters
    §§3.3 “From each window, we generate fixed-length sequences of eight consecutive quarters. Each sample is initially represented as a three-dimensional tensor X ∈ R^{T×N×F}, where the sequence length is T = 8.”
    provenance · paper-explicit
    stage · model fit
    12/74
  • B7
    rolling 28q window with 20 8 chronological split
    §§3.3 “For each eligible fund, we form overlapping windows of 28 quarters. ... Each 28-quarter window is then split chronologically into training and test subsamples, with the first 20 quarters reserved for training and the final eight quarters for testing. ... No random shuffling is employed, thereby preserving the time-series ordering of the data.”
    provenance · paper-explicit
    stage · model fit
    15/74
  • B8
    weighted categorical cross entropy default unit weights
    §§3.3 “Our estimation objective is a weighted categorical cross-entropy loss: L = -Σ w_c y_c log p_c ... By default, w_c = 1 for all classes. To enhance robustness to class imbalance and to emphasize hard-to-classify observations, we also implement an optional focal variant in which the loss is multiplied by (1-p_c)^γ, with γ=1.”
    provenance · paper-explicit
    stage · model fit
    61/74
  • B9
    min managers per stock filter
    provenance · benchmark-specified
    stage · signal gen
    72/74
  • B10
    out of sample predictions only for signal
    §§3.3 “We then compute model precisions for both the naive and LSTM-based predictors. We adjust our precision metrics by restricting attention to feasible cases. Specifically, we restrict attention to test outputs where the feasibility mask permits a sell decision.”
    provenance · paper-explicit
    stage · signal gen
    26/74
What this task reveals

The lowest-passing observed requirement is e2e metric consistency: 2 of 74 observed checker verdicts pass. The full property list above shows whether errors concentrate in domain-level validity or the paper-specific strategy logic.

Open paper → All tasks
Next in Factor Investing →
T12 · Factor Momentum and the Momentum Factor