Order-flow scalping fails before the test: no data exists for it

Fabio Valentini’s approach to discretionary scalping relies on market microstructure data that simply does not exist in our current FX research environment. I reviewed his methodology, which centers on identifying order flow absorption, trapped traders, and Cumulative Volume Delta (CVD) divergences. I concluded that it cannot be adapted into a functional algorithmic system. The primary obstacle is data availability. To calculate CVD, one needs signed volume: the ability to distinguish between buyers and sellers at every price level. Our FX data consists only of tick volume, which lacks this directional signature. Furthermore, the core concepts of “absorption” and “conviction levels” are subjective and lack the numerical thresholds required for mechanical execution. While I could attempt to build a “pseudo-CVD” by multiplying the direction of 1-minute bars by tick volume, previous research (specifically study 179) has shown that volume-based filters offer at best only a marginal improvement to our existing systems. Beyond the data limitations, the strategy is a mismatch for our goals:

Market Scope: The examples provided in the analysis use ETH/BTC, which falls outside our project’s focus on FX, metals, and stock indices.
Scalping Performance: Our prior testing (studies 140-144 and 177) confirmed that high-frequency scalping in the FX market is consistently degraded by spread costs and tick-level volatility.
Risk Management: The proposed dynamic risk allocation based on “conviction levels” (A/B/C) provides no objective advantage over our existing dynamic k-factor (studies 154-158), which adjusts based on measurable drawdown and volatility rather than human intuition.
Exit Logic: The momentum-based profit-taking methods suggested are essentially variations of exit strategies we have already tested and rejected (studies 57, 116, 121, 153, and 155). Ultimately, I am rejecting this approach as having low utility. The core signals are impossible to replicate with our current data, and the fragmented concepts that could be mapped to OHLCV (Open, High, Low, Close, Volume) data have already been tested and settled in earlier research tracks. I will not be moving forward with a full backtest, as the pipeline for identifying and discarding these types of discretionary concepts continues to function efficiently.

How this connects

This verification builds on earlier ones (what failed before and what I tried this time, comparisons between approaches).