What is the market maker model?
The market maker model describes price moving through a repeating multi-stage cycle — a consolidation where positions are said to be accumulated, a move that sweeps liquidity in one direction, a phase where the position is distributed, and a reversal into the opposite cycle. It is usually presented as a buy variant and a sell variant.
The framework sits alongside Power of Three and the broader smart money vocabulary, and shares most of its structural components: sweeps of prior extremes, displacement moves, and the return to an area before continuation. Where it differs is in the length of the cycle it describes and in the explicitness of the story about who is on the other side.
That story is the part that cannot be verified from price data. Whether a specific move was caused by a market maker filling an order, or by an ordinary imbalance between buyers and sellers, is not recoverable from a chart. Any test built on this model is therefore testing the geometry, not the explanation.
This is not a dismissal. A shape can be worth trading whether or not the story attached to it is accurate, and the shape is what a test can address.
Which parts of the model can actually be tested?
The testable components are the ones expressible as conditions on price: a sweep beyond a defined level, a displacement move of a stated normalised size, a retracement into a defined zone, and a continuation of stated magnitude within a stated time.
Each of those has an established treatment. The sweep is the same event covered in our liquidity sweep article. The displacement is a range-expansion condition. The retracement zone is usually a fair value gap or an order block, both of which have mechanical definitions. The continuation is an outcome measured at a fixed horizon.
Assembled, they form an ordinary multi-condition strategy: wait for A, then B within N bars, then C within M bars, then measure D. That is entirely gradeable, and the fact that it can be written this way is the test of whether a model has been specified or merely described.
What drops out in the translation is the narrative — the intent, the identity of the participants, and the claim that the cycle is engineered. None of it is needed to grade the rule, and none of it can be checked.
Why do multi-stage models look better than they test?
Multi-stage models look strong because each additional required stage filters the sample down to instances that already worked, and the filtering is invisible when examples are chosen by eye.
Consider a four-stage model. Looking for completed examples on a chart means looking for cases where all four stages occurred, which by construction excludes every case where the sequence began and broke down at stage two or three. The visible examples are the survivors of a filter nobody applied deliberately, and they will always look convincing.
The mechanical version has to count the failures. Every instance that reached stage one is in the denominator, whether or not it went on to stage two. Reported properly, a multi-stage model usually shows a large attrition at each step, and the honest result is the conversion rate at each stage alongside the outcome of the completed ones.
There is a second effect specific to long sequences. Each stage adds a parameter, and four stages with two or three thresholds each is a large parameter space. A model with that many degrees of freedom can be fitted to almost any history, which is why out-of-sample grading is not optional for this class of rule — it is the only thing standing between a model and a description of the past.
How do you avoid overfitting a multi-stage model?
Fix every threshold before grading, count how many combinations were examined, grade on a period the parameters never saw, and check whether the model beats a simpler version of itself.
The last check is the one most worth adding and least often performed. If a four-stage model does not outperform the two-stage version consisting only of the sweep and the continuation, the extra stages are consuming sample size without adding information. Comparing against the simpler rule on identical data is a cheap and unusually informative test.
Record the search. A model of this shape invites dozens of threshold combinations, and the best of dozens looks good even when nothing is there. The count belongs in your research log, and the article on backtesting journals covers why reconstructing it later does not work.
Walk-forward grading is the appropriate standard here rather than a single split, because it repeatedly re-fixes the parameters and grades forward, which is a closer analogue to how the rule would actually be used.
How does QuantParadox handle models of this shape?
QuantParadox grades multi-condition sequences mechanically, applies out-of-sample splitting by default, and reads the structural components as of the bar being decided rather than with hindsight.
The Chart Reader module is the relevant piece, because the components of this model are visual ones — structure, swings, ranges, compression, the shape of the candles. Reading them as of the decided bar is the specific defence against the retrospective labelling that makes multi-stage models look so much better by eye than they grade.
The Strategy Builder accepts the sequence described in plain English, which matters for a model normally held as a narrative rather than as code, and the Proof Engine grades the resulting decisions against history the strategy was never shown.
The honest expectation: rules with this many stages and thresholds fail out-of-sample grading often, and a platform that reports that failure is more useful than one that finds a parameter set which fits. Where a strategy does survive, the Reconciliation module identifies the conditions carrying it, which for long sequences is usually narrower than the model's description suggests.