Backtests are seductive. You build a strategy, run it over five years of historical data, watch the equity curve climb smoothly upward, and feel a quiet confidence that you have figured something out. Then you go live, and the market introduces itself properly.
Over the past couple of years, I have run three distinct algorithmic trading strategies in live markets. Not paper trading, not simulations. Real capital, real fills, real slippage. Each strategy taught me something different, and the lessons compound in ways I did not expect when I started.
The Gap Between Backtest and Live Is Always Larger Than You Think
Every practitioner knows about this gap in theory. Knowing it in theory and experiencing it are two entirely different things. The first strategy I ran was a mean reverting intraday system on index options. In backtests, the average fill assumption was clean. In live trading, the spread on far out-of-the-money strikes during the first thirty minutes of the session was wide enough to eat the entire expected edge on several trades.
The fix sounds obvious in hindsight: model your fills pessimistically. Add realistic slippage. Assume you will always buy the ask and sell the bid. But the subtler lesson was about which trades to skip. Not every signal the system generates is worth taking. Liquidity conditions at 9:20 AM are not the same as at 11:00 AM, and a strategy that does not account for that distinction will bleed slowly in ways that never show up cleanly in a backtest.
Position Sizing Is the Strategy
The second system I ran was a trend following approach on futures. The entry and exit logic was straightforward. What made the difference, in both directions, was position sizing. When I sized too aggressively relative to the strategy's historical drawdown, a routine losing streak felt catastrophic. Psychologically and financially. When I sized conservatively, the strategy performed almost exactly as modelled.
This taught me something that I think gets underweighted in most algo trading discussions: the position sizing model is not a secondary consideration, it is the core of the strategy. A mediocre entry signal with disciplined sizing will outperform a sharp entry signal with reckless sizing over any meaningful time horizon. Many investors and traders focus almost entirely on the signal generation side, and almost not at all on how much to risk per trade. That is backwards.
The practical framework I settled on was to size each position so that the maximum expected loss on a single trade, based on the strategy's historical worst-case, would not exceed a defined percentage of total capital. Simple, but it requires you to actually know your strategy's loss distribution, not just its average win rate.
Systems Break at Exactly the Wrong Moment
The third strategy involved a more complex execution setup, routing orders through an API to a broker. It worked smoothly for weeks. Then, on a day with elevated volatility, the API latency spiked. Orders that should have been placed in milliseconds took seconds. Some did not go through at all. The strategy's risk management logic, which depended on timely order placement, failed to function as designed.
This is the infrastructure lesson, and it is unglamorous but important. Your algo is only as reliable as the weakest link in its execution chain. That includes your broker's API uptime, your server's connectivity, your order management logic, and your fallback procedures when something goes wrong. I now build every strategy with an explicit "what happens if the exit order fails" scenario. Manual override procedures, alerts, position monitoring that is independent of the strategy itself.
The other thing this experience clarified: you need to be emotionally prepared to intervene manually when a system fails. Many people build algos partly to remove emotion from trading. But a failing system during a volatile session requires calm, fast, manual decision-making. That is a skill you should develop alongside the technical one.
Overfitting Is Subtle and Persistent
Across all three strategies, the most persistent challenge was overfitting. Not the obvious kind, where you have 47 parameters tuned to a 200-day backtest window. The subtle kind, where you make a series of individually reasonable-looking adjustments to a strategy based on recent live performance, and gradually sculpt it into something that fits the recent past rather than the underlying market logic.
One heuristic that helped: before changing any parameter in a live strategy, I ask whether the change is motivated by a genuine hypothesis about market structure, or by the fact that the last ten trades went against me. Most of the time, honestly, it is the latter. The discipline to leave a structurally sound strategy alone during a drawdown is harder than building the strategy in the first place.
Walk-forward testing helps, but it is not a complete solution. Markets regime-shift. A strategy that worked well in a range bound, low-volatility environment may simply not work in a trending, high-volatility one. Recognising when a strategy has stopped working versus when it is going through a normal drawdown is genuinely difficult, and anyone who claims to have a clean answer to that question is probably oversimplifying.
What I Would Tell Someone Starting Out
Start with a strategy simple enough that you can explain every trade it makes. Complexity in a live system creates blind spots. If you cannot explain why the system entered a particular trade, you will not be able to diagnose it when it starts losing.
Paper trading is useful for debugging execution logic. It is almost useless for understanding how you will behave when real money is at risk. The psychological dimension of live trading is not a minor footnote. It shapes every decision you make about when to intervene, when to trust the system, and when to pull the plug.
Finally, keep records of everything. Not just trade logs, but your reasoning at the time you made each decision about the strategy. When you look back six months later, you want to understand not just what happened, but what you were thinking. That record is where the real learning lives.
Running live algo strategies is genuinely one of the better ways to learn about markets, systems thinking, and your own decision-making under pressure. The tuition fee is real, but so is the education.