A Protocol-Frozen Evaluation of FLFT-Inspired Proxies: Penalization Usually Reduces the Feature Model to a Historical-Mean Baseline

Abstract Financial Liquidity Field Theory (FLFT) is a proposed framework. It treats money available for investment as a quantity that flows across a network of possible destinations, pushed by fundamentals, by collective mood, and by how easily capital can reach each destination. This paper reports a small study that does not test that framework. It reports the study instead as a demonstration of a research procedure. The test we wanted to run was flows first, prices second. We could not run it. We searched for a free source giving eight years of weekly, correctly time-stamped fund-flow history, and found none; no version of the regulatory filing Form N-PORT supplies it either. A rule fixed before we started then switched the study to a substitute question: do thirteen cheap, easily observed market indicators improve the ranking of next week's returns across twelve US-listed exchange-traded funds, compared with simple benchmarks? The important finding is about the model's tuning step, not about the indicators. Our main model, C04, uses a method that automatically shrinks weak indicators toward zero and can remove them entirely. It re-estimated itself 35 times over the protected test period. At 32 of those 35 re-estimations it chose the shrinkage setting λ = 0.10, and at that setting it set all thirteen indicator weights to exactly zero. We establish this by rebuilding the fingerprint that the frozen software stored for each set of weights, and matching it exactly to the fingerprint of a list of thirteen zeros. With all weights at zero, C04 becomes the historical-average benchmark C01. Its forecasts were therefore identical, digit for digit, at 126 of the 138 test weeks. They differed at 12 weeks and at 144 of 1,656 individual forecasts. Eleven of those twelve weeks changed the ranking score. Across the whole test the difference was exactly zero, and a decision rule fixed before we opened the test data returned MIXED_INCONCLUSIVE. Three counts must be kept apart. The main test sample is 138 weeks. C04 matched C01 at 126 of them. The indicators were actually in use at 12 weeks — and across those twelve, an exploratory range calculated afterwards spans about ±0.124 in ranking score, which is far too wide to learn anything from. A tuning step that sets weights to zero is itself a real out-of-sample result. But only twelve weeks tell us what happened when the indicators were used. The frozen evidence therefore does not show that these indicators carry no information in general. It shows that this particular tuned model usually discarded them, and added nothing to average ranking accuracy in this sample. No p-value appears anywhere in this paper. The plan was frozen and internally fingerprinted before the protected data were opened. No independent outside service timestamped it, so we do not use the word "preregistered". The contribution we defend is the procedure: a test set that can be opened only once, a marker recording that it was spent written before the first protected byte was read, a full public list of every trial run, a fixed decision rule, failed runs kept rather than deleted, and known defects disclosed. The negative result is the demonstration.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-14
DOI
https://doi.org/10.5281/zenodo.24220015
Primary Topic
Financial Markets and Investment Strategies
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

A Protocol-Frozen Evaluation of FLFT-Inspired Proxies: Penalization Usually Reduces the Feature Model to a Historical-Mean Baseline

Jasper Tristan Pong
Zenodo (CERN European Organization for Nuclear Research)
Financial Markets and Investment Strategies
article

A Protocol-Frozen Evaluation of FLFT-Inspired Proxies: Penalization Usually Reduces the Feature Model to a Historical-Mean Baseline

Jasper Tristan Pong
article en

Abstract

Abstract Financial Liquidity Field Theory (FLFT) is a proposed framework. It treats money available for investment as a quantity that flows across a network of possible destinations, pushed by fundamentals, by collective mood, and by how easily capital can reach each destination. This paper reports a small study that does not test that framework. It reports the study instead as a demonstration of a research procedure. The test we wanted to run was flows first, prices second. We could not run it. We searched for a free source giving eight years of weekly, correctly time-stamped fund-flow history, and found none; no version of the regulatory filing Form N-PORT supplies it either. A rule fixed before we started then switched the study to a substitute question: do thirteen cheap, easily observed market indicators improve the ranking of next week's returns across twelve US-listed exchange-traded funds, compared with simple benchmarks? The important finding is about the model's tuning step, not about the indicators. Our main model, C04, uses a method that automatically shrinks weak indicators toward zero and can remove them entirely. It re-estimated itself 35 times over the protected test period. At 32 of those 35 re-estimations it chose the shrinkage setting λ = 0.10, and at that setting it set all thirteen indicator weights to exactly zero. We establish this by rebuilding the fingerprint that the frozen software stored for each set of weights, and matching it exactly to the fingerprint of a list of thirteen zeros. With all weights at zero, C04 becomes the historical-average benchmark C01. Its forecasts were therefore identical, digit for digit, at 126 of the 138 test weeks. They differed at 12 weeks and at 144 of 1,656 individual forecasts. Eleven of those twelve weeks changed the ranking score. Across the whole test the difference was exactly zero, and a decision rule fixed before we opened the test data returned MIXED_INCONCLUSIVE. Three counts must be kept apart. The main test sample is 138 weeks. C04 matched C01 at 126 of them. The indicators were actually in use at 12 weeks — and across those twelve, an exploratory range calculated afterwards spans about ±0.124 in ranking score, which is far too wide to learn anything from. A tuning step that sets weights to zero is itself a real out-of-sample result. But only twelve weeks tell us what happened when the indicators were used. The frozen evidence therefore does not show that these indicators carry no information in general. It shows that this particular tuned model usually discarded them, and added nothing to average ranking accuracy in this sample. No p-value appears anywhere in this paper. The plan was frozen and internally fingerprinted before the protected data were opened. No independent outside service timestamped it, so we do not use the word "preregistered". The contribution we defend is the procedure: a test set that can be opened only once, a marker recording that it was spent written before the first protected byte was read, a full public list of every trial run, a fixed decision rule, failed runs kept rather than deleted, and known defects disclosed. The negative result is the demonstration.

Zenodo (CERN European Organization for Nuclear Research)
Financial Markets and Investment Strategies
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon