SPY Averaging Ladder

The ladder buys SPY at many nearby prices during the session and sells the cheapest lots as soon as the market lets them clear a small, fixed gain. Each buy is small. The strategy relies on SPY moving back and forth more often than it moves in one direction.

• Over 445 backtested sessions, from 2 January 2025 to 12 October 2026, the profit-taking engine realised +$54,062 after commissions. It was positive in every one of the 22 months.
• The loss is concentrated. On the 25% of sessions where SPY's range from 04:00 to the close exceeded 1.5%, the ladder lost −$19,057. The remaining 75% of sessions made +$10,161.
• The latest changes reduced the loss but did not remove it. Trading from 04:00 and widening the step as the day moves improved the net result from −$13,232 (regular hours, fixed step) to −$8,896. Widening the step also halved the sessions that used all 30 lots, from 43% to 21%.
• The current settings were chosen on 2 October 2026 using data up to 1 September. Over the 28 sessions since, which played no part in choosing them, the ladder took +$2,421 in profit and finished −$200 net, in line with the earlier results.
• The strategy is in research. Current work focuses on standing aside, or holding fewer shares, on trending days before any capital is committed.

The same six rules run on every price update. They are applied in this order, and the live system and the backtest use the same code.

1. 01

   Set the day open and reference

   At startup the system loads the one-minute bars since midnight New York time. The first bar's open, normally 04:00 premarket, is the day open. The first live quote becomes the reference. The previous day's close is never used.

2. 02

   Step the grid

   Whenever the price moves one step from the reference, up or down, the reference moves to the new price and a 10-share buy is triggered. The step is 0.05%, plus 0.01% for every 0.1% SPY has moved from the day open in either direction.

3. 03

   Hold one lot per level

   A buy is skipped if a lot is already held within one step of the price. Selling a lot frees its level. No more than 30 buys are placed per session.

4. 04

   Sell in pairs

   The two cheapest lots are sold together once the pair shows at least +0.10% on its cost after all four commissions: two buys and two sells.

5. 05

   Sell only at the floor or better

   Every sell is a limit order at the lowest price that still meets the target, so a stale quote cannot turn a planned gain into a loss.

6. 06

   Pause, then resume

   After a sale the system waits 10 seconds before buying again, so it does not immediately buy back the level it just sold.

SPY's day open, the 04:00 premarket bar, was $699.50, and it is now trading at $700.00. Every price below stays within 0.1% of the day open, so the step stays at its base of 0.05%, which is $0.35 at this price. Commission follows the Interactive Brokers fixed tier: $0.005 per share with a $1.00 minimum, so each 10-share order costs $1.00.

At $700.00 the pair would have made only +0.097% after costs, so the ladder waits for $700.03. Lots A and D stay open and follow the same rule as a pair. The four $1.00 minimum commissions use up almost a third of the target on a trade this size, which is why the target is always measured after costs. Had SPY been 0.1% or more from its day open, the step would have been 0.06%, and buys would have come $0.42 apart rather than $0.35.

Premarket and regular hours, from 04:00 New York time to the regular close, 2 January 2025 to 12 October 2026. Settings as in the term sheet above, including the widening step. Each session starts with no position, and any lots unsold at the close are valued at the closing price.

Sessions tested

4452 Jan 2025 – 12 Oct 2026

Profit taken

+$54,062After $14,886 in commissions

Positive sessions

67.4%300 of 445

Largest drawdown

−$10,215Peak to trough, cumulative net

Results by intraday range

Intraday range is the session's high minus its low from 04:00 to the close, as a percentage of the close. The median session moved 1.05%.

Configurations tested

All rows cover the same 445 sessions with 10-share lots, a 30-lot limit and a +0.10% pair target. "Widen 0.01%" adds 0.01% to the step for every 0.1% SPY moves from its day open; "down only" widens only while SPY is below it. "Limit reached" is the share of sessions in which all 30 lots were bought. Trading from 04:00 adds profit taken without a matching rise in losses at the close. From 04:00, widening the step mainly reduces risk: it lowers the worst sessions and the drawdown and holds fewer lots on trending days, while improving the net result by up to $800. The 0.005% setting has the best net result, but the margin over 0.01% is under $1 per session, too small to tell apart from chance.

A backtest follows one price history. Two standard quantitative tools can show how much the result depends on that history. They can also show whether the sessions that cause the losses can be spotted early enough to act. The figures below come from the 445 backtested sessions of the current setup. They show what each tool can do, not a finished model.

Markov chain: do volatile sessions cluster?

Each session is classed as calm, normal or volatile by its intraday range. A Markov chain estimates the probability of each class tomorrow given the class today. In the table, each row is the previous session and the shading shows probability.

• Intraday regime detection

  A hidden Markov model on one-minute returns and the running range would estimate, through the day, the probability that the session is trending. The ladder would stop buying or reduce its lot limit once that probability passes a threshold.

• Regime-based lot limit

  The prior-day class, which does carry information about tomorrow's range, would set the starting lot limit: 30 after a calm day and fewer after a volatile one.

• Grid as a Markov chain

  Each grid step is treated as a move up or down with probabilities estimated per regime. This gives the expected time to clear a pair and the probability of reaching the 30-lot limit, which can be checked against the backtest.

• Regime-aware settings

  The step already widens as SPY moves away from its day open. A regime model could also set the base step, the widening rate and the profit target before the session starts, based on the expected regime.

Monte Carlo: how likely is the result?

Monte Carlo simulation builds many possible trading years and runs the strategy through each, so a setting is judged on a distribution of outcomes rather than one history. As a first pass, the backtested sessions were resampled into 20,000 simulated years of 252 sessions. Sessions were drawn in blocks of five, so volatile spells stay together.

Median year

−$4,863Current settings

Bad year, 5th percentile

−$12,5891 year in 20 is worse

Good year, 95th percentile

+$2,1571 year in 20 is better

Years with a loss

87.0%Of 20,000 simulated

Median drawdown

−$7,230Peak to trough within a year

Deep drawdown, 5th pct.

−$13,974Guide for capital buffers

• Synthetic price paths

  One-minute SPY paths would be generated from the regime model above, and the strategy's code would run on them, so settings are tested on many more sessions than history provides.

• Parameter robustness

  Grid step, target and lot limit would be compared by their full outcome distributions. This favours settings that hold up across paths over settings that fit one history.

• Stress tests

  Extra trend days, gap opens, and spread and slippage costs would be added to measure how far results worsen outside the backtest's assumptions.

• Capital and sizing

  The drawdown distribution would set the capital held against the 300-share maximum position and the point at which trading is suspended.

How the two fit together

1. 01

   Estimate regimes

   A Markov model is fitted to one-minute SPY data, with transition probabilities between calm, normal and trending states.

2. 02

   Simulate sessions

   Monte Carlo generates thousands of price paths that switch between those states with realistic frequency and persistence.

3. 03

   Run the strategy

   The same code that trades live runs on every path, with and without the regime filter, under the same costs.

4. 04

   Compare distributions

   A change is adopted only if it improves the median year and the bad-year and drawdown tails together.

The sessions that cost the ladder money are usually trending sessions, and trends are often started by news: an inflation print, a Federal Reserve decision, or results from one of the index's largest companies. Price-based models, including the Markov model in section 6, detect a trend only after it has begun. Language models can read news and official text as it is published and score its tone and relevance. That could give an earlier warning, or one that does not depend on price. This section describes how sentiment could be used. It has not yet been tested.

What would be read

How it could steer the ladder

• Pre-session gate

  The overnight score and the day's event calendar would set the starting lot limit. The ladder would trade with fewer lots, or not at all, on sessions with major releases or an unusually negative news backdrop.

• Intraday brake

  A sharp shift in headline tone during the session would pause new buys. Buying would resume after a set period without further shifts. Existing lots would still be sold at their normal targets.

• Direction bias

  The ladder can already be set to buy on dips only, on rallies only, or both. Strongly negative news could switch it to rallies only, so that it stops buying into a news-driven decline.

• Input to the regime model

  Sentiment scores would be added as an input to the Markov model in section 6. The probability of moving into a trending state would then depend on the news as well as on price.

How it would be tested

1. 01

   Build a timed archive

   Headlines, releases and transcripts would be collected for every backtested session, each stamped with the minute it was first published.

2. 02

   Score without hindsight

   Each item would be scored using only text available at that minute, with a fixed model version so scores can be reproduced.

3. 03

   Test the signal

   Scores would be checked against each session's range and direction, and in particular against the sessions with a range above 1.5%.

4. 04

   Test the strategy

   The ladder would run with the sentiment rules through the backtest and the Monte Carlo simulation. A rule is adopted only if the whole distribution improves.

Limits

• Look-ahead bias. A language model trained on data after a test date may already know what happened. Backtests would use models whose training data ends before the period being tested, or keep the most recent period for testing only.
• False alarms have a cost. Sessions up to a 1.5% range, 75% of the total, made money. A gate that closes too often would give up that profit to avoid the losing days.
• Speed. Markets react to headlines in seconds. The ladder works on minutes, so sentiment is better suited to limiting risk than to trading on the news itself.
• Data and model dependence. News feeds and model access add running costs and outside dependencies. Their output can also change when a vendor updates a model, which is why model versions would be fixed.

• Daily buy limit

  No more than 30 buys per session. This caps what can be held at 300 shares.

• Sell floor

  Every sale is a limit order at or above the price that meets the target after commissions. No sale is sent at a price below it.

• Average-cost guard

  Sales are also held above the account's average cost for the whole position, so the broker never records a sale as a realised loss.

• Stale-price guard

  The system does not trade on the previous close. If only delayed data is available, it warns the operator, and its sell limits keep delayed quotes from causing losses.

• Extended-hours orders

  Outside regular hours, buys are limit orders at the ask. Market orders are never used outside regular hours.

• Startup reconciliation

  On restart, the system rebuilds its lots from seven days of executions and cancels any orders it did not place, so it does not act on an outdated position.

• Widening step

  As SPY moves away from its day open, buys are spaced further apart. On a trending day the 30 lots cover a wider price range, and fewer of them are bought.

• Day-open fallback

  If the bars since midnight are not available, the first live price of the day becomes the day open. The day open resets at midnight New York time.